Tanmay Khanna, Lucknow, India
Building intelligent systems across data, AI, agents, automation, and open source.
I work at the intersection of the layers that have to hold together before an AI system is useful. Pick a stage below to find it on the map.
I like to understand a system from first principles before I add anything to it. That habit shows up everywhere in my work: reading a codebase until its structure makes sense, breaking a vague problem into constraints, and turning a complicated idea into software that runs.
So far my work spans backend systems, AI applications, RAG, agents, MCP, multimodal data, automation, data processing, open source, and evaluation. It moves between low-level engineering, like deduplicating audio in bulk, and higher-level AI design, like deciding how an agent should behave when a fixed workflow breaks.
I'm early in my career. My current role at InfoBay.AI began in 2026, so what I bring is depth in a few areas, breadth across the stack, and a steady rate of learning. The rest of this page shows the work instead of claiming it.
whoami: Tanmay Khanna, AI Systems Engineer. Focus: multimodal data, agent architectures, evaluation. Status: early career, learning fast, building in public.
InfoBay.AI
I work across the data lifecycle required to make multimodal AI systems possible.
The job is not scraping. It's everything between raw files and a dataset a model can be trained or evaluated on: deciding what belongs, structuring it, checking it, removing what shouldn't be there, and moving it safely at scale. I've worked with audio, images, video, and multimodal datasets.
Each step exists to make the next one trustworthy.
Exact hashes only catch byte-identical files. Similarity-based detection is how near-duplicates get found.
Complex and multilayered images, including:
These formats carry structure, such as layers, vectors, and scenes, that ordinary image pipelines flatten or drop.
AI-generated video detection connects to my research project, SAGA.
DICOM files carry rich metadata, which makes filtering and PII handling part of the job.
A model is only as good as what it was trained and tested on. My work at InfoBay.AI exposed me to the practical machinery around that: where data comes from, how it moves, what gets rejected, and how the survivors are stored. That is the part of AI that rarely shows up in demos.
Select a stage to see what happens there.
Two systems I'm designing and building. Each diagram is interactive: hover a technology to see where it sits, or run the walkthrough.
Architecture
A retrieval and reasoning service built around a planner, hybrid retrieval, reranking, and a reasoning step that produces the final response. Conversational memory and evaluation are part of the design, and I'm using it as a place to experiment with LoRA and QLoRA.
Select a stage for details.
Illustrative trace. No model is running on this page.
In progress
A privacy-oriented, local multi-agent system that automates parts of the engineering job-application workflow. It's a personal systems project and an experiment in local, privacy-conscious, resilient browser automation. It is not a finished production system.
Use deterministic automation whenever the workflow is known, and use agents only when deterministic automation cannot reliably handle the situation.
agentic fallback, used only on a miss
who is acting
Walkthrough of the intended flow. Nothing is executing.
| Metric | What it tells me | Status |
|---|---|---|
| Application completion rate | How often a started application reaches a valid submission. | Implementation in progress |
| Human intervention rate | How often a person has to step in. | Implementation in progress |
| Workflow recovery rate | How often a failed step recovers without help. | Implementation in progress |
| Model calls per application | The cost of using agents at all. | Implementation in progress |
| Latency | Time from posting to ready-for-approval. | Implementation in progress |
| VRAM usage | Peak memory on the 12 GB target. | Implementation in progress |
| Deterministic vs agentic ratio | How much work the cached workflows absorb. | Implementation in progress |
No numbers appear here because none have been measured yet.
Research / Exploration Phase 1 implemented: real vs AI Source attribution: open
A research and engineering project about identifying AI-generated video and investigating where it came from. It combines video analysis, spatial and temporal transformations, and questions of provenance and dataset quality.
Phase 1 has been implemented for real vs AI detection. Source attribution remains an active research challenge, particularly because reliable attribution depends heavily on appropriately labeled and detailed training and evaluation data.
Select a stage for details.
Sensor noise, compression, and stabilisation leave traces on genuine footage. A detector has to separate "unusual" from "generated", and those are not the same thing.
Low-quality CGI can share visual tells with generated video. The question is about origin, not realism, which makes it harder than it sounds.
When artists use AI inside a traditional pipeline, it's no longer clear what "AI-generated" should mean. The label itself becomes ambiguous.
Generated footage composited with conventional graphics blurs the source signal. One video can have several origins at once.
Re-encoding, cropping, and re-uploading strip metadata and degrade the traces that attribution relies on.
Flagging real footage as AI has a real cost. It needs to be measured and analysed on its own, not averaged away inside one headline number.
Attribution needs labels that say which generator, which version, and what pipeline. Detailed labels like that are scarce and hard to verify.
No accuracy figure appears on this page. I'd rather show the failure modes than a headline metric.
Evolving direction
This is a direction I'm growing into, not a credential. I'm interested in how models behave, where they break, and how to measure that honestly. Topics I'm reading about and experimenting with:
model
Illustrative example of the method, not a reported result.
Contribution
Caravan is an open-source Bitcoin multisig coordinator. I worked in its monorepo on maintainability and the interface.
My fork (opens in a new tab) Upstream repository (opens in a new tab)
Where I worked in the codebase. Select the root to expand it.
Two overlapping representations consolidated into one shared type, so changes happen in one place.
ESLint moved from v7 to v9, including the configuration changes that come with it.
Layout improvements so the interface works on smaller screens, in React with Tailwind.
Smaller, clearer structures that are easier for the next contributor to read.
Exploration / Active work
This is exploration, not a finished contribution. I'm not claiming any merged work here.
Public
An Express.js server scaffolding library, run as a CLI. It sets up a production-shaped backend in one command: project structure, dependencies, database wiring, and an optional AI module.
$ npx create-jerry
Structure as documented in the project README.
Public
A full-stack AI services platform with a separate frontend and backend, user authentication, and persistence.
TypeScript
A live-streaming platform with Lightning and Bitcoin payments, using QR codes and payment receipts.
Bitcoin concepts built from scratch in Python: how Bitcoin works, how Bitcoin Script works, and a take on the transaction optimization problem.
Docker, Kubernetes, Argo CD
A DevOps project that containerises a simple Express app, deploys it to a Kubernetes cluster with Argo CD, and manages a canary release with Argo Rollouts. The README documents each step and the difficulties I hit.
Not ranked. Depth varies: some of these I use constantly, some I've used on a project or two, and a few I'm still learning.
My GitHub profile is the primary proof of work. Anything on this page that is public links to a repository you can open. Work at InfoBay.AI is private, and for projects above without a repository link, the write-up on this page is what is public so far.
$ gh repo list TomLucasakaTGeek
Notes from building, written in public: what I built, what broke, and what I measured. These are the entries planned so far. None are written yet.
Themes I plan to write about: AI systems, agent architectures, RAG, evaluation, open source, system design, multimodal datasets, engineering experiments, and failure analysis.
If you're working on data infrastructure, agent systems, evaluation, or open source, and this page suggests I could be useful, I'd like to hear from you.