Orbweaver · Kadi · KrishiMitra · Vayu · MargaDrishti · VaidyaMitra · AGENTIQ
I build AI systems for problems that already have victims: crop loss, air quality, road capacity, fraud rings, misdiagnosis, thin-file credit. And I build them so a sceptical reader can check every claim I make.
I work across the whole range rather than one corner of it:
| What I reach for | Where it shows up | |
|---|---|---|
| Classical ML | XGBoost · LightGBM · CatBoost · scikit-learn · igraph | Orbweaver's account scorer, KrishiMitra's crop model, CreditSetu's risk tiering |
| Deep learning | PyTorch · ResNet50 / VGG16 / EfficientNetV2 · CNN-LSTM · GraphSAGE · Grad-CAM | OpenForensics' three-backbone ensemble, the retinopathy grader, Vayu's forecaster |
| Generative AI | Gemini · Amazon Bedrock (Nova Pro) · RAG · vision OCR | VaidyaMitra reads strips and reports, Kadi's grounded bilingual assistant, Specledger's extraction |
| Agentic systems | LangGraph · MCP tool layers · planner + executor splits | SmartAlloc's 7-agent pipeline, AGENTIQ's permission-checked tool layer, Inflx |
The thing that stays constant across all four is not the technique.
A model is allowed to propose. Something deterministic, a threshold, a knapsack, a peeling objective, an assertion evaluator, is what decides. That separation is the single design decision I repeat most, because it is what makes "why did this happen?" answerable by a person.
flowchart LR
E["evidence in"] --> D["deterministic<br/>parse · validate · features"]
D --> M["<b>the model proposes</b><br/>XGBoost · CNN · LLM · agent"]
M --> G{"calibrated,<br/>enough evidence?"}
G -->|no| A["<b>abstain</b><br/>route to a human"]
G -->|yes| DEC["<b>deterministic decides</b><br/>peeling · knapsack · assertions"]
DEC --> O["output + what it cost<br/>evidence · ₹ · false positives"]
classDef learned fill:#151B2E,stroke:#818CF8,stroke-width:2px,color:#EDF4FF
classDef proved fill:#0E2230,stroke:#22D3EE,stroke-width:2px,color:#EDF4FF
classDef plain fill:#0F1422,stroke:#64748B,color:#EDF4FF
classDef soft fill:#101A2B,stroke:#38A8FF,color:#EDF4FF
class M learned
class DEC proved
class E,D,O plain
class A,G soft
Why it is worth the extra work. Ring membership in Orbweaver comes from a peeling objective with a proved ½-approximation bound, so "why is this account in this ring?" is checkable arithmetic rather than a model's opinion. AGENTIQ generates test assertions with an LLM and then evaluates them with a tool, because a model grading its own output is not evidence. Specledger's extraction works with the LLM switched off entirely. The model adds recall, it is not load-bearing.
| Project | The hard part |
|---|---|
| KrishiMitra | CatBoost crop recommendations cross-checked against five years of government district returns, leaf disease at 93.75% over 10,162 images, FAO-56 irrigation advisory, Soil Health Card baselines from 13.35M tests. 12 languages, deployed free |
| Vayu | LightGBM + CNN-LSTM forecasting over a 15,360-cell satellite grid, Gaussian-plume ROI ranking, and difference-in-differences verification that an intervention actually worked, never a guessed AQI |
| Kadi | 59,985 siloed FIRs into one explainable link graph across 31 districts and 298 stations. Shared modus operandi ranks as a hypothesis, never as a name. The translator refuses to touch FIR numbers, dates and identifiers |
| VaidyaMitra ⟨org⟩ | Every identifier is masked before it reaches the model. Jan Aushadhi generic matching with substitution-safety warnings, vision OCR, ten Indian languages, on Amazon Bedrock |
| MargaDrishti | Bengaluru road-capacity loss on one H3 × hourly substrate: 298k violations, 8 model families, and a published audit of enforcement bias in its own training data |
| Floodcast-Gurugram | Will this route flood, and when. Live rainfall matched against 73 researched flood points, rather than a city-wide alert nobody can act on. Provenance is a first-class field: 39 of 73 points carry a named source and every row has a confidence value |
| Diabetic-Retinopathy-Detection | Temperature-scaled confidence with reliability diagrams and ECE; low-confidence cases escalate to a human grader. Front page says not a medical device, because it is not |
| Medicure-AI | Photograph a strip → composition, NPPA price, Jan Aushadhi generic, interaction warnings, each with a calibrated confidence and an honest refusal when evidence is thin |
| CreditSetu | pip install creditsetu. Validated against 150,000 real borrowers with real default outcomes: 0.82 AUC using only 7 of 14 features, to close the circularity gap of testing on its own synthetic data |
| Project | The hard part |
|---|---|
| Orbweaver | Densest-subgraph extraction over a 35.7M-edge account graph: 0.7292 ring precision against a 0.2242 base rate, always reported with the 0.371 real customers swept in per fraudster caught. 36 dated failures published alongside |
| AGENTIQ | B.Tech final-year project. Eight vulnerability families mapped to the OWASP API Security Top 10, probed by baseline differential, so a finding needs a material deviation rather than a suspicious-looking string. Every outbound request passes a permission-checked, SSRF-guarded, audited tool layer |
| artifact-repro-triage | Checks whether a paper's repository contains what its README promises. 0% → 100% detection of fabricated file claims across 742 artifacts |
| OpenForensics | Three-backbone deepfake ensemble with calibrated confidence and per-backbone Grad-CAM. The dashboard shows the evidence, not just the verdict |
| Specledger | A logistic calibrator over 11 evidence features picks an auto-publish threshold hitting a measured precision floor on held-out data, instead of trusting an LLM's self-reported confidence |
| MedGuardX | Context-aware PII/PHI masking: an engine on PyPI, a hardened FastAPI service with JWT RBAC, and an app on top |
| Project | The hard part |
|---|---|
| SmartAlloc | A 7-agent LangGraph pipeline over linear programming that finds compute waste and predicts SLA bottlenecks before they land |
| Talent-Intelligence | 100,000 candidates ranked in under 18 seconds, CPU only, with honeypot and fake-profile filtering |
| CacheLLM | pip install cachellm-proxy. A drop-in proxy keeping the OpenAI request and response shape including streaming, verified against the official Python and Node SDKs. Exact hash, then a cosine tier whose safe threshold is measured per embedding model. 77% hit rate, 78% lower cost, zero wrong answers on genuinely new questions |
| Adaptive-Graph-Search-Suite | Graph traversal on realistic map topologies, built to be watched while it runs |
The claim is cheap; these are the times it cost me something.
| Where | What I published anyway |
|---|---|
| MargaDrishti | A target of PR-AUC ≥ 0.45 was set assuming ~10% prevalence. The real label rate is 0.291%, so the goal was unreachable by construction. Reported as a 46.9× lift over base rate with the original goal marked wrong, not as a 3× shortfall |
| MargaDrishti | Seven model families all returned PR-AUC 0.9999 on one task. That is the signature of a recovered business rule, not a hard problem, so it is reported as a recovered rule, because presenting it as modelling performance would mislead |
| MargaDrishti | The review process changed regime mid-window, so every model on that task is miscalibrated. Reported as not-yet-answerable rather than as a weak result |
| Orbweaver | Four of thirteen investigations came back negative and are published beside the nine that worked, including one where the hypothesis was exactly backwards |
| CreditSetu | The live demo runs on synthetic data, and the README says so above the numbers rather than below them |
| HackerRank Orchestrate | 24-hour agentic AI challenges: #12 (Sep 2026), #290 with a bronze medal, top 15% (Aug 2026), #427, top 25% (Jun 2026). Each submission is public: Buy or Wait, Message Notification Router, Multi-Modal Evidence Review |
| KSP Datathon 2026 | Reached Level 2 with Kadi, built end to end over 50 days and deployed for the Karnataka State Police |
| Hacksplosion 2026 (Deloitte India) | Cleared Levels 1, 2 and 3 |
| SWITCH Energy-X | 11th of 93 on the private leaderboard, recovering a hidden physical law from 500,000 unlabeled, 26%-missing sensor rows |
| micro1 Frontier Engineering | artifact-repro-triage, measured across 742 repositories |
| PyPI | Three packages published and documented: cachellm-proxy, creditsetu, medguardx-core |
10+ national hackathons and datathons, each shipping something that runs.
Shipped on FastAPI, Streamlit and Next.js; deployed to Cloud Run, Vercel,
Render and AWS; packaged to PyPI where it makes sense. But the stack matters
less than the discipline around it: a temporal split a test enforces, a
held-out set nothing touches, the false-positive cost printed next to the
detection rate, and a FAILURES.md recording what I got wrong on the way.
The banner, the timeline and PROJECTS.md are generated from the GitHub API, my own repositories and both organisations, so none of them can quietly fall behind what I have actually shipped. The three avatars (mine and the two organisations') are drawn by the same scripts, in one visual language:
make refresh # pull the current repository list
make assets # redraw the banner and the timeline
make index # rewrite PROJECTS.md
make # the last twoClaiming reproducibility on 32 projects and then hand-maintaining my own profile would have made this the one dishonest page on the account.
Final-year computer science at LNMIIT Jaipur. I build under two organisations, VaidyaMitra for clinical work and B-TechProject for my final-year project. Most of what I build ends up deployed somewhere free, because a model nobody can open is a claim nobody can check.
