139
automated tests
43 contract, 62 backend, 10 IoT, 17 ML and 7 frontend tests, plus a 41-scenario live smoke test.
PharmaChain
Tanmay 2.0 · Applied AI engineering
The power of AI.The mind behind it.
AI isn't the achievement. What I build, verify and ship with it is.
I work with AI agents the way a good engineer uses any powerful tool: clear requirements, small steps that can be checked, and proof that the result works. These six builds show how, including what went wrong along the way.
- 6
- Projects built with AI agents
- 139
- Automated tests in one project
- 2,510
- Cross-document facts checked by rules
- 3%
- AI claims caught by my validation gate
How I work with AI agents
AI does the typing. I own the thinking.
A repeatable loop that keeps speed and quality together. It is visible in the repositories: written architecture notes, phased builds and recorded trade-offs.
- 01
Frame
I define the problem, constraints and what "done" means before any code is written. Larger builds start with an architecture document and a living status file the agent can read.
Evidence architecture.md + context.md in PharmaChain and FactMesh; PROJECT_DECISIONS.md in RepoLens
- 02
Delegate
Agents plan and implement in small phases I can review. I pick the model for the job: a stronger one for architecture, a faster one for implementation.
Evidence PharmaChain: 7 phases · RepoLens: phases 1 to 6 · FactMesh: separate planning and implementation models
- 03
Verify
Nothing counts as done until it has run against real data. I use tests, ground-truth fixtures and live end-to-end runs, and I fix what the actual output shows.
Evidence 139 tests + a 41-scenario live smoke test · a fixture with expected results for scoring accuracy
- 04
Decide and document
I own the trade-offs and write down what does not work yet. That is what lets someone else trust, extend or challenge the result.
Evidence Decision logs, known-limitations sections and "what I would build next" lists in every major repo
Impact, not hype
Numbers I can point to.
No inflated claims. Each figure below is taken from the project documentation, with the project it comes from.
2,510
facts cross-checked
Across 6 PDFs and 511 pages. Each verdict keeps a saved rule trace, so it can be explained without calling a model.
FactMesh
3%
AI claims rejected
104 of 3,302 extracted claims were thrown away because the quoted evidence was not on the page.
FactMesh
0.94
anomaly precision
Recall 0.76, F1 0.84, on a held-out synthetic test set. Labelled as synthetic because no real counterfeit data exists.
PharmaChain
11
endpoints with ground truth
A bundled test service with a written list of expected results, so accuracy is scored instead of eyeballed.
RepoLens
0
network requests
Your saved profile never leaves the browser, and the extension never submits a form for you.
FillMate
Where human judgment stays in charge
- Letting rules decide verdicts while the model only explains them, so results are reproducible.
- Safety limits: never executing untrusted code, never submitting a form, never guessing a field.
- Honest scope: a local chain only, a modular monolith instead of microservices, SQLite instead of a graph database.
- Naming the gaps: synthetic training data, synonym blindness in entity states, demo-grade authentication.
Projects
Six builds. Real problems. Checked results.
Each project shows the problem, what stands out, how it was verified and the judgment calls I made along the way.
FactMesh
Finds where documents agree, disagree, or only look like they disagree.
The problem
Analysts comparing filings and economic reports cannot easily tell a real contradiction from a difference in time period, scope or units.
What stands out
- Rules decide, AI explains: every verdict can be reproduced from a saved rule trace.
- Grounding gate: every extracted fact must quote the page word for word, or it is discarded.
- Model layer that falls back automatically from Gemini to a local Ollama model when free-tier quotas run out.
PharmaChain
Fake-medicine detection and drug traceability on blockchain.
The problem
Counterfeit drugs enter supply chains because no single party can be fully trusted. A pharmacist or patient needs one clear answer when they scan a pack.
What stands out
- Scan a QR code and get one verdict: genuine, suspect, counterfeit, recalled or expired.
- Machine-learning anomaly model combined with rule checks such as impossible travel speed and scan bursts.
- Offline-capable checks: the Merkle proof travels inside the QR code, and the interface says clearly what still needs a live check.
RepoLens
Paste a GitHub URL and get an API map, an OpenAPI spec and load-test results.
The problem
Understanding and stress-testing an unfamiliar Spring Boot service takes hours of reading code and wiring up tools by hand.
What stands out
- Reads the Java source without running it, so untrusted repositories are never executed.
- Finds every endpoint and parameter, estimates which need authentication, and gives a confidence level rather than a guarantee.
- Load tests report real latency percentiles (p50, p90, p95, p99) and throughput, not just an average.
SmartPark
A parking platform that allocates slots using weather, demand and emergencies.
The problem
Drivers lose time searching for a slot, and operators have no live view of load. Emergency vehicles also need priority.
What stands out
- Weather-aware slot allocation, with priority for ambulances, fire trucks and police vehicles.
- Floor scoring by occupancy, exit distance and user preference, plus predictions from past occupancy logs.
- QR gate entry, indoor navigation, a booking-expiry worker and an admin dashboard for peak hours and zone load.
FillMate
Fills application forms from your saved profile in one click.
The problem
Filling in the same personal, education and experience details on every application form is slow and repetitive.
What stands out
- Reads each field’s label, placeholder, name and autocomplete hint, then matches it using a synonym dictionary.
- Built to be trusted: it never submits, never overwrites what you typed, and leaves uncertain fields blank.
- Works in Chrome and Brave, with a test form covering many labelling styles.
RTI Portal
A digital workflow for filing and tracking Right to Information requests.
The problem
Filing an RTI request and following its status is slow and opaque for citizens and for the officers who respond.
What stands out
- Three roles: citizens file and track requests, officers respond against deadlines, admins see analytics.
- Request timeline, personal dashboards and a production database schema alongside the demo.
- Responsive, dependency-free front end that opens straight from a file.
My AI toolkit
The right tool for each job.
I choose models by what the task needs, and I keep a fallback so the work does not stop when one option fails.
Planning, building and debugging
Claude
A stronger Opus model for architecture and planning, and a faster Sonnet model for implementation. I work in a pair-programming style and fix things based on the real output.
Extraction at scale
Gemini
Used for fact extraction in FactMesh. Its free tier limits each model separately, so I built an adapter that moves down a list of models automatically.
Fallback and privacy
Ollama (local models)
A local model that needs no API key. It took over when every hosted option ran out for the day, and it is a real fallback, not a theoretical one.
Working practices that make agents reliable
- Written architecture and status files so every session starts with context.
- Phased plans with a test or check at the end of each phase.
- Grounding gates and rule engines around model output.
- Provider-agnostic adapters, so no single vendor is a point of failure.
Experiments and failures
What went wrong, and what I did about it.
Working with AI quickly means finding failures quickly. These are real ones from the projects above.
FactMesh
Two-column pages read as one
On a two-column report, text from the left and right columns was joined mid-sentence, which produced nonsense facts.
What I did Added column detection. A wide table that spans both columns is still mis-ordered, and that limit is documented.
FactMesh
A safety gate that was too strict
The quote check correctly rejected a model answer that was not word for word, but it also rejected a real figure computed from two true numbers on the page.
What I did Kept the gate and wrote down the trade-off between strict grounding and recall. Next step: a "derived fact" type that stores both source quotes.
FactMesh
The free quota ran out mid-build
Each Gemini model had its own daily limit, as low as 20 requests, and I hit all of them.
What I did Built an automatic fallback across models and then to a local model. The final showcase explanations were generated that way.
PharmaChain
An anomaly model with no real fraud data
No real counterfeit-drug data exists, so a supervised model would have been dressed-up guesswork.
What I did Trained an unsupervised model on synthetic normal patterns, reported its numbers as synthetic, and added rule checks alongside it.
Let's talk
Want an engineer who moves fast and still checks the work?
I adapt quickly, pick up new tools as they appear, and care about proving that things work. If that is what your team needs, let us talk.
← Back to the main portfolio© 2026 Tanmay Kaushik