Srinivas Sivaratri
Practical AI systems, developer tools, and open-source contributions — built close to the metal and useful in the real world.
I like software that can explain itself: the small tools that make a system safer, the strange experiments that make an idea tangible, and the contributions that hold up under someone else's review.
This is a working archive of that practice.
Available — open to remote full-time or contract engineering roles Based in India · UTC+5:30 (IST) · comfortable overlapping APAC and European time zones
Select a capability to jump to its closest body of work.
Signing in with an unregistered email and a compromised password returned PASSWORD_COMPROMISED instead of the generic error, revealing which emails were registered. Traced it to the haveIBeenPwned plugin throwing during the dummy-password hash, and proposed a request-scoped skip flag plus a regression test.
Status: maintainers adopted the approach and shipped it in #5998.
geth db truncate-freezerA subcommand for archive nodes that binary-searches the merge block, preserves headers and hashes, and drops pre-merge bodies. Reviewed across three rounds by core maintainers (s1na, jwasinger, MariusVanDerWijden); revised it to batch-process rather than hold ~15GB in memory on mainnet, to move headers into a fresh freezer instead of assuming equal table lengths, and to look up merge blocks per network.
Status: core team built their own version of the command.
Built ETL pipelines. Sped up DBs. Automated ops: -50hrs/mo.
I start by naming the domain before I touch code, then keep that language in a maintained CONTEXT.md so the product and implementation do not drift apart. I grill the idea adversarially until the weak assumptions and unsafe edges are visible enough to cut or test. From there I break the work into vertical slices, each small enough to ship and prove end to end. Every slice gets a diff review and an explainer quiz, so I can defend not only what changed but why the design still holds. The result is a process that uses AI for speed without outsourcing judgment.
Like receiving the insurer's rejection letter three weeks before it exists — every red line becomes one physical ask: a document to request or a question for the doctor.
Selected builder — one of ~150 chosen from 5,000+ applications (≈3%) for the five-hour Bengaluru build sprint.
POSTDATED turns photos of hospital discharge paperwork and bills into a predicted, itemised insurance rejection while the paperwork can still be fixed. It separates money already lost from recoverable gaps, then translates each recoverable line into something the family can physically ask for at the counter.
TypeScript, Next.js 16, React 19, Tailwind CSS 4, the Anthropic SDK, and Vitest. The hard part is separating multimodal document extraction from deterministic policy arithmetic and a negation-aware guard that blocks unsupported medical claims. The money and guard paths have 38 tests, and phone photos are resized to a 2,576px long edge before upload.
Like writing a secret in the sand while waves wash it away, then seeing if an AI can reconstruct your original words from the remaining dots.
Bun and TypeScript, with a Python/FastAPI helper and Hugging Face inference. The hard part is degrading text in real time while capturing newly restored words at their exact fuzz level without sending the original text to the server. Each round makes at most one model call; only the final fuzz and fresh clues cross the client boundary.
Like opening Schrödinger's famous box, where a split-second AI calculation decides if you find a cozy cat or an empty space.
Python, PyTorch, Diffusers, and Stable Diffusion Turbo. The hard part is capturing and serving real intermediate denoising frames while leaving outcome selection authoritative on the backend. Both outcome branches use a 512px, six-step image-to-image run, and only the selected branch is generated.
Like having a local shopping assistant read through hundreds of product reviews online, summarize them, and explain the verdict out loud in your regional language.
Next.js, TypeScript, Neon Postgres, Upstash Redis, Sarvam AI, Gemini, and Firecrawl. The hard part is carrying voice input through evidence-gated review synthesis, localization, and text-to-speech while streaming partial results to the browser. It supports 11 languages and caches reviews and localized results for 30 days.
Like an AI puzzle master solving Sokoban warehouse games by starting from random movements and gradually refining them into a perfect path to the solution.
Python, PyTorch, NumPy, and Flask. The hard part is a conditional diffusion policy combining an 8×8 CNN board encoder with a Transformer action denoiser trained on BFS-derived solutions. It samples four 20-action candidates in 10 DDIM steps and replans for at most 20 iterations.
Like a virtual coding assistant working inside a designated project folder, writing, testing, and fixing program loops automatically until they work.
Python, Gemini 2.0 Flash, google-genai, and python-dotenv. The hard part is a function-calling loop that returns filesystem and execution results to the model while injecting a working-directory boundary. It exposes four tools, caps the agent at 20 model turns, and limits Python execution to 30 seconds.
Like describing a scent you love to someone who has already smelled every perfume in existence, watching them pull every bottle that shares its character from memory.
Like a movie critic who understands exactly what you mean when you describe a film scene, even if you can't remember the title or actors.
Like a digital book club where margins come to life, allowing readers to share thoughts and discuss stories in real-time as they read.
Like a digital accountant scanning invoice PDFs to extract the total amount while double-checking its own math to ensure it never makes a mistake.
Like a terminal-based morning newspaper aggregator that pulls articles from your favorite blogs and stores them in a local library database.