I'm Kshitij Sarve, a full-stack AI engineer. I turn voice agents, retrieval systems and document workflows into products, from the first API to the interface and cloud.
From AI products to GPU physics: the scope of recent builds, followed by measured improvements from earlier roles.
Scale, workflow complexity & verification
investors in the matching pool
Fundraising · MyProBuddyLayered retrieval before AI writes the report.
Explore the platformstages in the lending workflow
Underwriting · MyProBuddyDeterministic rules. Cited explanations. Human review.
Explore the workflowphysics tests checking the renderer
Horizon · Personal projectGPU checked against CPU. Reported 170M rays/s.
Explore the rendererBefore → after · earlier roles
faster video pipeline
Inagiffy · GenAI videoAlso cut cost per video: $3.00 → $0.50.
cheaper sentiment calls
Inagiffy · Sentiment analysistranscription accuracy
Quikscribe · LLM transcriptionUnderstand the workflow before choosing a model or writing an API.
Balance answer quality, latency and cost. Keep human review where it matters.
Connect the interface, backend and cloud, then evaluate what ships.
The problem, the system and the decisions behind it. Client projects are anonymised.
A fundraising workflow that brings voice onboarding, document extraction and investor matching into one platform.
Key decisionMatch in layers: SQL filters narrow ~47K investors to a few thousand, vector search picks the top 30, and only then does the model write the report.
A multi-tenant lending platform that checks documents, applies lender rules and prepares evidence-backed assessments for human review.
Key decisionCode computes every number and rule; the model only explains them, must cite evidence, and can never override a failed check.
Reads long procurement documents, checks eligibility and prepares a grounded first draft.
Key decisionThe agent drafts; a person signs off before anything is sent.
Turns patient context and medical guidelines into recommendations with traceable sources.
Key decisionNo claim without a citation, checked by LLM evals before it ships.
A GPU renderer that traces light around a spinning black hole. A physics experiment built for fun, pair-programming with an AI coding agent.
Key decisionTrust, but verify: the analytic shadow outline is drawn over the ray-traced image, and 68 physics tests check the GPU against the CPU.
Watch animation ↗Agents, voice & grounded answers
From the API to the interface
Retrieval & reliable data
Ship, operate & protect
A growing scope of responsibility, one role at a time.
to
AI Developer Intern
Candidate matching & grounded answers.
Screening time halvedShipped a candidate matcher and a grounded assistant for internal finance documents.
to
Freelance Backend Developer
A backend for meeting transcription.
Accuracy 75% → 95%Built the backend for LLM meeting transcription.
to
AI Developer · Product
Faster video. Cheaper sentiment calls.
Renders 5 min → 30 sCut cost per video from $3 to $0.50 and made sentiment calls 99% cheaper.
Now to present
AI Backend Engineer
Fundraising & lending. Frontend to cloud.
2 AI platforms, end to endLead engineer on two AI platforms, end to end: a fundraising platform with voice agents and vector investor matching, and a lending platform with AI document checks and underwriting. Frontend, backend and cloud.
Engineer by trade. Here's the rest of me.
Working remotelyI like problems that reward curiosity, in code, in physics and in the markets.
Outside work, I trade ideas with other builders and follow interesting questions into side projects. Horizon started with one: what would light look like around a spinning black hole?
Indian indices & F&O. A daily lesson in risk and probability.
Trading ideas with ambitious builders.
Learning by building. Interesting ideas deserve to run.
Good engineering means knowing what to automate, what to measure, and when a person should decide.
An AI product, an engineering role, or a good conversation. Say hello.
