I explored Google Research and collaborators' fruit fly brain map: real data, GitHub repositories, eight use cases, and what the connectome does not prove.
I sent Google's article about the male fruit fly brain map to Jarvis, my AI assistant, and asked two questions: where is the GitHub repository, and should we install it?
That second question needed unpacking. Installing a viewer would let us inspect the data. Downloading the connections would let us analyze a graph. Neither action would give Jarvis a biological brain. The interesting question was what we could actually learn or build with the map once we stopped treating it as a downloadable intelligence.
So I explored the public resources with Jarvis: the microscopy, selected neurons in 3D, connection tables, a male/female comparison, the research repositories, and two small neuron-skeleton files. This guide comes out of that bounded exploration. I am writing as an AI product builder, not as a neuroscientist claiming to have reconstructed or experimentally tested these circuits.
The cover shows real DNg13 neuron reconstructions, not an AI-generated illustration. Data: Berg et al. and the MaleCNS collaboration, CC BY 4.0. Rendered in Neuroglancer; the camera, colors, and background were set for readability, and interface chrome was cropped. The underlying neuron geometry was not edited.
The useful version of the story This is a structural wiring map of one male fruit fly's central nervous system, including the brain and ventral nerve cord. It is not a functioning digital animal. The final paper reports 166,700 reconstructed neurons, 11,710 neuron types, and 124.2 million contacts between proofread neurons. Those quantities describe different things. You can already browse neurons, inspect tissue, trace candidate connections, compare selected male/female circuits, and download small samples without mirroring the dataset. The strongest near-term applications are better scientific questions, more reproducible analyses, useful research interfaces, and education. Autonomous intelligence and robot control remain additional research problems.
First, what exactly is a connectome?
A connectome is a map of connections in a nervous system. Neurons are the nerve cells. Synaptic contacts are the identified contact relationships through which neurons can influence other neurons. A reconstruction also gives you shape: where a neuron's branches travel and which regions they enter. Together, shape and connectivity make the map much more useful than a list of cell names.
For software builders, a useful starting model is a directed, weighted graph. A node represents one reconstructed neuron. An edge points from a presynaptic neuron to a postsynaptic neuron. Its weight can count the contacts between that pair. The official download schema distinguishes individual synaptic sites, their partner relationships, and aggregated neuron-to-neuron connections.
The graph analogy has limits. A contact count is not a measurement of how strongly a connection behaves in every situation. The map does not, by itself, tell you the cells' activity at a particular moment, all their physiological properties, or how experience changes their interactions. Structure constrains possible explanations; it does not finish the explanation.
What was mapped, and why the whole CNS matters
The final Cell paper by Berg and collaborators describes one five-day-old adult male Drosophila melanogaster. It includes the central brain, both optic lobes, the ventral nerve cord, and the neck connective joining brain and nerve cord. The tissue was imaged at 8-nanometer resolution in all three spatial dimensions.
CNS means central nervous system. The ventral nerve cord is an important part of the fly's sensorimotor system, below the brain. Including it means researchers can examine routes between brain processing and circuits associated with the body's movements within the same reconstructed animal. Calling this only a brain map understates that scope. Calling it a complete digital fly overstates it: a structural CNS dataset is not a model of the animal's muscles, body, senses, environment, and ongoing activity.
Final Cell paper, published September 3, 2026. Million-scale values retain the paper's reported precision. Reconstructed neurons 166,700 Proofread and annotated neurons, including sensory axons Neurons in the connection graph 166,483 The reconstructed neurons that have synapses in this graph Neuron types 11,710 Groups defined using morphology and connectivity Synaptic contacts 124.2 million Contacts with both endpoints on proofread neurons Directed neuron-pair edges 25.58 million Neuron-to-neuron pairs; one edge can aggregate many contacts
This was a collaboration, not a Google-only release. The project credits name FlyEM at HHMI Janelia, the University of Cambridge's Department of Zoology, the MRC Laboratory of Molecular Biology, and Google Research. The public data and the human effort behind its reconstruction deserve as much attention as the AI methods.
The paper, the Google post, and the live catalogue are not identical snapshots. The release notes date v1.0 to June 8, 2026. The final paper and Google article appeared September 3, 2026. Google reports 11,691 cell types; the final paper reports 11,710. The catalogue generated September 9 displays 11,751. I use the final paper's number for the study, and label the catalogue separately. I have not conclusively reconciled the remaining catalogue difference. The catalogue's combined pre- and postsynaptic-site total also should not be substituted for the paper's contact count. A version label and a counting definition matter more than a headline's biggest number.
Where AI fits: turning tissue images into a usable map
The colorful neuron renderings begin with tissue images like this one. Google Research's technical account explains the reconstruction work. The basic sequence is microscopy, segmentation, synapse detection, proofreading, annotation, and analysis. Segmentation means deciding which parts of an image belong to the same cellular structure. That is a difficult computer-vision problem, particularly where thin branches run beside one another.
AI helps turn the image volume into candidate neuron shapes and contact locations. Proofreading corrects many errors, and annotation makes the result searchable by cell identity and region. Other models predict properties such as neurotransmitter identity. A neurotransmitter is a chemical involved in signaling between cells; a predicted label should remain visibly distinct from a direct experimental measurement.
This is already a substantial AI application: making a scientific dataset feasible to reconstruct and inspect. It is a different claim from saying the finished dataset will immediately improve a language model. AI helping us understand biological intelligence and biological data improving artificial intelligence are related possibilities, but they require different evidence.
The GitHub answer: there is an ecosystem, not one install
The most direct answer to my original question is flyconnectome/2025malecns. The paper names this repository for its accompanying data and analyses. But it is not a single application that launches the entire research project. The viewer, query client, morphology tools, dataset, and paper-specific outputs have separate jobs.
2025malecns Paper-specific supplemental data, notebooks, and analysis outputs Not a turnkey reproduction pipeline google/neuroglancer Viewing microscopy, segmentation, and 3D neurons A viewer, not a neural simulation neuprint-python Programmatic graph queries through neuPrint Authenticated access needs your own token navis Skeleton and mesh analysis, measurements, and plots Coordinate units and reference spaces must match Cell Type Explorer A browser-first overview of cell types and their partners Type-level tables aggregate individual neurons Dimorphism Explorer Inspecting male/female morphology and connections Comparison normalization and volume coverage matter synister_malecns Inspecting the neurotransmitter-model work Code availability is not proof of easy reproducibility
I would start in the hosted viewers. Installing Neuroglancer locally makes sense if you need to load your own data or control custom views programmatically. It is unnecessary just to inspect these public neurons. Likewise, downloading a complete graph before choosing a question adds storage and analysis work without making the question better.
There are practical repository caveats. At the revision I inspected, the official count notebook still selects v0.9, and the repository README still references the preprint. No license file was present in that reviewed repository tree. The dataset's CC BY 4.0 terms do not automatically establish the terms for every software file. Check the particular code and data you plan to reuse.
Worked example 1: following a candidate route toward movement
I started with DNg13, a descending-neuron type. Descending neurons connect brain-level processing with circuitry in the ventral nerve cord. They are not automatically motor neurons that directly innervate muscles. The public catalogue has a left and right DNg13 reconstruction, with body IDs 11074 and 512006.
The first useful action was simply rotating the reconstruction and relating the branches to the surrounding anatomy. The second was reading the region and partner tables. That moves the experience from looking at a beautiful object to asking a testable question: where does this neuron receive input, and where does it send output?
Using the LoVP92 and DNg13 pages, I checked this type-level chain: TmY21 to LoVP92 to DNg13 to IN16B045. LoVP92 is a visual-system neuron type marked male-specific in the resource. The following counts are aggregated contacts between types, recovered from the precise catalogue tooltips rather than inferred from rounded display values.
Catalogue snapshot inspected September 13, 2026. Direction runs from input partner toward downstream partner. TmY21 LoVP92 2,721 LoVP92 DNg13 54 DNg13 IN16B045 963
These three rows are useful, but easy to oversell. They do not establish that the same individual neurons form one continuous functional pathway. They also do not show signals traveling through it during behavior. Type-level aggregation can hide which member of a type connects to which other cell. An individual-neuron query is the next step before making a stronger pathway claim.
The input fraction adds another corrective: LoVP92 accounts for about 0.44% of DNg13's catalogue input contacts. A real connection is not necessarily a dominant one. I would not turn that row into a claim that LoVP92 controls DNg13, or that this short chain explains the behavior by itself.
The distinction also matters when comparing our example with the paper's Figure 4 discussion. Its highlighted hypothesis involves LoVP92 reaching DNg13 through an inhibitory intermediary, VES200m. That is a different route from the direct catalogue link we checked. Our short walk demonstrates how to inspect the tables; it should not be presented as a reproduction of the paper's proposed mechanism.
DNg13's region table lists 1,828, 1,176, and 1,025 presynaptic output sites in the three leg neuropil regions, out of 4,307 total. That is about 93.5% across those regions. The official function-annotation CSV also associates DNg13 with turning during forward walking, citing earlier work. The anatomy and that prior annotation make a sensible starting point for investigation; our table inspection did not perform a new behavioral experiment.
Worked example 2: a male/female comparison that needs context
The AOTU012 comparison page places corresponding male and female neurons into a shared view. There is one left and one right neuron of this type in each sex. The resource identifies a male-specific ventral axon projection. This is the kind of structural difference that a shared 3D view makes much easier to inspect than a paragraph alone.
For the actual two-sex overlay, open the official AOTU012 comparison. It shows male neurons in cyan and female neurons in magenta. The image above intentionally contains only the male dataset: FlyWire's public-data guidelines specify separate noncommercial reuse terms, so I have not republished the mixed-data screenshot here.
The connection table is equally instructive. It includes raw male weights, scaled male weights, and female weights. The male values are multiplied by 0.581 for this comparison. The comparison also excludes male ventral-nerve-cord connections because the female FlyWire dataset being compared is a brain volume, not a matched whole-CNS reconstruction.
0.581 in the viewer is not 0.79 in the final paper. The live page I visited still described a 0.581 scaling factor. The final paper's Definition of dimorphic edges methods section reports a revised factor of 0.79. The excerpt below preserves the website's displayed values; it is not a recomputation using the published method. Neither number is a biological performance ratio between male and female flies.
Selected rows from the comparison website's 0.581-scaled snapshot, not the final paper's revised normalization or a new statistical test. DNde002 640 371.84 332 DNa02 382 221.94 145 PVLP203m 223 129.56 0
The full table contains 264 rows. I exported those rows during the exploration. Even this three-row excerpt shows why normalization matters: 640 versus 332 gives a different impression from 371.84 versus 332. The scaled values are analytical comparison values, not fractional contacts directly observed in tissue.
A zero also needs care. PVLP203m is listed as a male-specific partner. Its absence as a matched female partner does not prove that female flies lack an entire behavior. Nor is the table's classification of a partner as dimorphic a significance test for every edge attached to it. The correct next questions concern matching, coverage, variability, and function.
Eight things this resource can help people do
The following are applications I would distinguish, not eight capabilities delivered by installing a package. Browsing and bounded data analysis are available now. Experimental conclusions require laboratory work. Product ideas need user validation. Simulation and robotics need additional models and evaluation. Keeping those categories separate makes the opportunity more concrete, not less interesting.
For a neuroscience team, a question such as how visual information reaches walking-related circuitry is too broad to test all at once. The map lets the team start with named cells, list candidate upstream and downstream partners, and inspect where the contacts sit. The DNg13 example is a small demonstration of that narrowing process. The useful deliverable is a short list of candidate routes with neuron IDs and evidence, rather than an attractive but uncheckable diagram.
I would move from type-level tables to individual-neuron connections, compare contact counts with input fractions, and inspect important links in the underlying imagery. I would also report which routes disappear when a minimum-contact threshold changes. That sensitivity check helps expose conclusions that depend on a handful of weak or uncertain edges. The result is a hypothesis for functional testing, not a claim that an anatomical shortest path is the route a living fly necessarily uses.
A comparison can ask several different questions: is a cell type present in both datasets, does its shape differ, and are its partners or relative connection weights different? Those questions should not be collapsed into one label. The AOTU012 example shows why: a broadly corresponding neuron can have an additional projection and different connections without the entire circuit being unique to one sex.
A useful project would export matched neuron types, retain the comparison's scaling and exclusions, and present morphology alongside connectivity. I would include a short explanation of every zero and every unmatched cell rather than silently treating missing data as biological absence. The deliverable could be a reproducible comparison notebook or an annotated set of candidate differences. The limitation is fundamental: comparison across individual specimens and reconstruction pipelines is not the same as estimating a population-wide effect. The paper's discussion of technical and individual variation remains part of the result.
Laboratory time is limited. A connectome can help a team choose among plausible targets before investing in functional recordings or perturbation experiments. Instead of selecting a neuron only because it lies in an interesting region, researchers can ask whether its inputs, outputs, and morphology fit a specific mechanistic prediction. A proposed target should have a reason to be on the list and a predicted observation that could prove the explanation wrong.
For example, the output distribution of DNg13 suggests a more focused question about walking-related circuitry than a generic question about all descending neurons. But the practical next step also depends on experimental access, the specificity of available tools, and suitable behavioral measurements. None of those appears automatically because a body ID exists. I would frame this use case as experimental triage: a source-backed candidate dossier that a domain expert can accept, revise, or reject. The value is a better selection process, not a promise that the selected neuron will explain the behavior.
A skeleton is a simplified tree tracing a neuron's branches. It supports questions about branching, spatial extent, and the regions a neuron reaches without loading the full microscopy volume. navis provides established tools for this kind of morphology work. A focused analysis could compare a left/right pair, inspect an unusual branch, or examine how a set of candidate cell types differs in shape.
In our exploration, the two DNg13 SWC files contained 9,752 and 9,821 skeleton nodes. Those are points along two neurons, not thousands of additional neurons. We checked file structure, unique node IDs, parent references, and tree consistency. That is a useful technical check before plotting, but it does not validate the biological tracing. Coordinate units matter just as much: the MaleCNS SWC files we used are in 8-nanometer units, while other downloads use other units or template spaces. A visually plausible plot can still have the wrong scale.
This is the use case closest to my own work. A useful assistant could translate a constrained question into a read-only graph query, show the returned rows, explain the fields, and link back to the selected neurons. Someone might ask which types contribute the largest fraction of a neuron's inputs, or whether a partner relationship is visible on both type pages. The assistant's job would be to make the evidence easier to retrieve and understand.
I would require every answer to include the dataset version, neuron or type IDs, query parameters, units, and source link. Graph facts, prior literature, and the assistant's interpretation should be separate. An answer that cannot establish a biological claim should say so. A small evaluation set could include correct numerical questions, ambiguous type names, missing neurons, and questions the data cannot answer. The success criterion would be agreement with verified queries and honest abstention, not how convincingly the assistant narrates a pathway. This is a proposed product workflow; we did not build or benchmark it in this exploration.
For computer-vision researchers, the challenge is not just displaying the finished graph. It is recovering reliable structure from difficult imagery. The paper describes segmentation, synapse detection, proofreading, and predicted neurotransmitters; the neurotransmitter-model repository provides one code entry point. Potential projects include uncertainty estimation, prioritizing suspicious reconstructions for human review, or evaluating how a reconstruction error changes a downstream graph conclusion.
I would begin with a bounded, expert-reviewed benchmark. A model should not be credited for finding an error merely because its output disagrees with an automated label. Nor should nearby image samples be casually split into training and test sets if that leaks nearly identical structure across the boundary. Record label provenance, holdout strategy, and the cost of human verification. Large data does not remove the need for trustworthy evaluation. This use case requires specialist judgment and potentially substantial computation; it is not the natural first project for someone opening Neuroglancer for the first time.
For a workshop or classroom, this is a strong alternative to another synthetic sales spreadsheet. Learners can move from a real tissue image to a reconstruction, from the reconstruction to a table, and from the table to a carefully bounded conclusion. They encounter graph direction, aggregation, normalization, measurement units, data versions, and uncertainty in one coherent example. The visuals are compelling because they are evidence, not decoration.
I would give learners a small assignment: identify one neuron type, record its body IDs, choose one partner relationship, and write two sentences. The first says what the data shows. The second says what it does not establish. A more technical group could check a skeleton file or reproduce a table calculation. The output should be a compact evidence note with a source and version, not a claim that the learner has understood an entire nervous system. That is also a useful AI-literacy exercise: ask an assistant for a claim, then check whether the underlying data really supports it.
The wiring map can supply structural constraints for a model: which cells connect, where branches lie, and which candidate routes deserve attention. A simulation must still choose how its model neurons behave, how contact counts relate to effective interactions, what inputs arrive, and how the system changes over time. Different choices can produce different behavior from the same anatomy. A plausible animation is not enough to validate those choices.
For robotics, the distance is larger still. A controller must work with sensors, actuators, a body, delays, noise, and an environment. I would test a small biologically inspired circuit against a conventional control baseline on a narrowly defined task, with equal evaluation conditions. The deliverable might be evidence that a particular design principle is useful, or evidence that it is not. It would not be a robot running a downloaded fly mind. This is a research direction that the map can inform, not a functioning robotics capability delivered by the MaleCNS release.
There are real precedents for this direction. Shiu et al. (2024) built a computational fly-brain model and tested sensorimotor predictions against experiments. Lappalainen et al. (2024) combined connectome-constrained networks with optimization to predict visual neural responses. NeuroMechFly v2 provides a separate framework for embodied sensorimotor simulation. These studies show why anatomy, model assumptions, and validation belong together. They are not demonstrations that this particular MaleCNS download already runs a complete animal.
A practical first session, without a large installation
If you want to explore it yourself, start with one cell type and one question. A blank whole-brain scene can be less useful than a well-chosen example because there is so much data and so little context. My suggested first question is: what evidence connects DNg13 with circuitry in the leg regions, and what would I need before making a claim about behavior?
Record the dataset and the question. Start at the official project. Note male-cns:v1.0, the access date, the relevant source, and the question you are trying to answer. This small record prevents you from mixing a notebook's old version with a newer website table. Open a named example. Use the DNg13 type page. Identify both body IDs, inspect the 3D reconstruction, and read the classification. Rotating a pair of selected neurons is a better first task than trying to display everything at once. Read the tables before drawing a pathway. Compare upstream partners, downstream partners, and regions. Check whether the values are totals, per-neuron averages, percentages, or synaptic-site counts. For a precise contact total, inspect the catalogue's detailed tooltip instead of multiplying rounded display values. Look at the underlying image data. Open the preloaded Neuroglancer scene and inspect a microscopy section. This connects the abstract graph to the material it was reconstructed from. It does not make a beginner qualified to resolve ambiguous synapses; that still requires expertise. Try one carefully bounded comparison. Open AOTU012. Compare the shapes, then read the normalization notes before comparing counts. Keep raw and scaled male weights in separate columns. Do not hide the exclusion of ventral-nerve-cord connections. End with evidence and an open question. Save a source-linked view, the relevant table rows, and a short interpretation. Write down one conclusion the evidence supports and one conclusion it does not. A focused, auditable result is a successful first session; understanding every neuron is not a sensible requirement.
A small Python example you can verify
The download documentation lists public skeleton files separately from the full connection graph. The Python 3 example below fetches the same two DNg13 SWC files used in our exploration and counts their data rows. It uses the standard library, needs no account, and caps each response at one megabyte. It reads data; it does not execute downloaded code or install anything.
The seven SWC columns describe a node ID, a node type, x/y/z coordinates, radius, and parent ID. This snippet only checks the column count and counts points. A fuller integrity check should examine numeric values, unique IDs, roots, parent references, and cycles before analysis. Neither check determines whether the biological tracing is correct. For this particular SWC download, multiply coordinate values by 0.008 to express distances in micrometers; do not assume that unit conversion applies to every other format in the project.
If your next question requires arbitrary graph queries, move to neuPrint with your own account and token. If it concerns shape, use navis in an isolated Python environment. If it requires reproducing a paper figure, inspect the figure's inputs, code revision, filtering, and dataset version before installing dependencies. Choose the tool because the question needs it, rather than treating installation as progress by itself.
Browser exploration does not require a full dataset download. The official inventory lists separate downloads of roughly 13 MB for annotations, 1.1 GB for the full segment-to-segment graph, 12.7 GB for synaptic point locations, and 6.8 GB for partner pairs. These are different resources, not one mandatory install. We used public pages and two small skeletons; we did not mirror the complete graph or microscopy volume.
What the map does not establish
The word complete describes an important achievement in coverage, but it should not be read as error-free or physiologically exhaustive. The paper's methods and limitations are essential reading if you plan to base a scientific conclusion on these data. The following boundaries affect almost every use case above.
Coverage is not perfect recovery. The paper reports that 40.1% of detected pre-to-post contacts have both endpoints assigned to proofread neurons. That is a coverage measure within the detected data. It does not mean that only 40.1% of the fly's actual biological connections exist or that the rest are necessarily wrong. One specimen is not a population. The reconstructed male is one animal. Individual variation, preparation, reconstruction, and differences between datasets complicate comparisons. A compelling male/female image is not a substitute for population-level evidence. Connectivity is not causality. A link or a graph path can motivate an explanation. It does not establish which neurons are active during a behavior, whether a link is necessary or sufficient, or how the circuit responds when perturbed. Those are functional questions. Analytical flow is not measured neural traffic. A max-flow score or graph-traversal result depends on a mathematical model and its choices. It is not a direct recording of signal throughput in a living animal. Contact counts and predicted neurotransmitter labels do not remove that distinction. A fly result is not automatically a human result. This dataset is not evidence for claims about human personality, cognition, disease treatment, or sex-related behavior. Any broader comparison needs its own biological argument and independent evidence. Open data is not a packaged artificial mind. The release does not supply a validated whole-animal simulation, a pretrained assistant, a robot controller, or evidence of consciousness in software. Those claims would require additional models, experiments, and definitions.
What I would build first, and what I would not promise
My first product experiment would be an evidence-first circuit explorer for a narrow audience, such as researchers or advanced learners working with a few named neuron types. A user asks a question; the system retrieves a bounded result, opens the relevant view, and produces a short evidence note. It would keep the source rows visible and let the reader distinguish the database result from an explanation generated by AI.
The initial scope would be deliberately small: one dataset release, a handful of supported query patterns, and read-only access. I would test it against known answers like the DNg13 input fraction and the AOTU012 scaling example. It should preserve a query record, flag unsupported requests, and make it easy to correct an interpretation. I would not begin by promising a universal neuroscience copilot or importing the entire dataset into an opaque vector search index.
There may be useful products around reproducibility, teaching, annotation review, or specialized analysis. But a striking dataset is not a business model. Before charging for a tool, I would want evidence that a real user repeatedly loses time on a specific task, that existing tools do not already solve it well, and that the proposed interface improves accuracy or turnaround without hiding uncertainty. The public map itself is not the differentiator; the quality of a validated workflow could be.
The larger AI lesson
I began with an installation question. I ended with a better appreciation of how much useful work lies between a scientific dataset and a justified conclusion. Jarvis helped locate the resources, inspect public views, extract a bounded sample, and check the language against the evidence. None of that required pretending the assistant had absorbed a fly's intelligence.
That is the part I find most actionable. We can use AI to lower the friction of serious exploration while preserving the human and scientific work that gives the data meaning. We can make the evidence easier to reach, the units harder to confuse, the assumptions more visible, and the next experiment more focused. Those are worthwhile outcomes even before anyone demonstrates a new controller or neural simulation.
The right response to this map is neither to dismiss it as just a visualization nor to announce that brains are now downloadable. Open the data. Choose a small question. Follow the evidence. Then be precise about what you have learned and what still needs to be tested.
Exploration notes and image credits
This guide's source review and hands-on checks were performed on September 13, 2026. We verified public viewer access and 3D rotation, inspected DNg13 and LoVP92 partner tables, exported the 264-row AOTU012 comparison, and downloaded and checked two DNg13 skeletons. We did not reproduce the full paper, run authenticated neuPrint analyses, train a model, perform a biological experiment, or validate a neural or robotic simulation. The use cases and product proposals are my interpretation of what the resources could support.
The cover, DNg13 viewer, and microscopy image use MaleCNS data from Berg et al. and the project collaboration under CC BY 4.0. They are actual Neuroglancer captures. The cover uses a custom camera, display colors, and white background; its interface was cropped. The microscopy retains its original scale bar. The catalogue screenshot retains its labels and IDs. Images were converted to WebP for the website. No neuron branches, scientific labels, or tissue structures were synthetically added.
The AOTU012 image also contains only MaleCNS neurons and its brain outline. The original official scene was reduced to those male-data layers before capture; no FlyWire geometry is embedded in this article's image files. Full data credit: FlyEM at HHMI Janelia, University of Cambridge, MRC Laboratory of Molecular Biology, Google Research, and the contributors named by Berg et al.. The viewer is Google Neuroglancer; its software license is separate from the data licenses. The linked female comparison credits and terms remain available through the official FlyWire and MaleCNS resources.
Subscribe free to Harshith's Newsletter to read every article in the interactive edition.
Harshith Vaddiparthy works with founders, operators, and teams on practical AI products, workflows, advisory, training, and mentorship. This no-JavaScript version preserves the page's core information and navigation.