Suvrakamal DasAI engineer at MattoboardKolkataIST

I'm Subhro, a software engineer building AI agents: the skills they learn, the orchestration that runs them, and the infra underneath.

Merged PRs in PyTorch and vLLM, spoke at PyCon and MCP Dev Summit, published at SciPy. Here's my journey in a nutshell, or find me on X.

Subhro at the Golden Gate Bridge
Golden Gate, San Francisco

Shanghai, where I last spoke. Home, Kolkata, is at the bottom of the page. From a photo by King of Hearts, CC BY-SA 4.0

Main Quests

  1. MattoboardMachine Learning EngineerOct 2025 – Present
    Remote
    • Product Finder — built the pipeline that takes a room photo and returns the exact materials in it: fine-tuned YOLO segmentation, per-segment metadata extraction, then hybrid search (vector similarity + metadata filters with fallback) over an optimized product table, replacing LLM re-ranking at about $0.02 per material.
    • Similarity Search API — a "more like this" API on Postgres with pgvector HNSW that returns similar materials and products in under 300 ms worldwide, with post-search diversification (over-fetch top 100, group by parent product) so results aren't just color variants.
    • Image Upscaler — fine-tuned a GAN-based upscaler on 30K+ Topaz-upscaled images to fix the grainy texture that builds up after 3–5 AI edits, replacing a third-party API costing about $0.10 per image across a 1M+ image pipeline.
    Show 3 moreShow less
    • Segmentation Model — built the labeled dataset and trained a custom YOLO segmentation model that isolates chairs, curtains, floors and other items in room photos, handling edge cases general-purpose models miss.
    • Model Inference — deployed the YOLO and GAN models for GPU inference on Modal.
    • Ownership — worked directly under the CTO with no manager above me, owning features from design to production, including UI changes.
  2. XRI Global AIAI EngineerOct 2024 – Apr 2025
    Tennessee, USA · Remote
    • Dashboard & Data Integration — built a full-stack language technology dashboard for African language data (BLEU/chrF++ charts, Leaflet.js maps, REST APIs, PySpark), improving data accessibility by 40% with automated database backups.
    • LLM Trans-tokenization — developed a cross-lingual vocabulary transfer strategy to adapt Mistral-7B to new languages, reaching 87% accuracy on Hindi–English translation with better BLEU, perplexity and chrF than the base models. Built token alignment with SMT tools and fine-tuned on low-resource languages across Latin and non-Latin scripts.
  3. ProjectX.Cloud (YC P26)Machine Learning EngineerOct 2022 – Oct 2024
    Hyderabad, India
    • The product — ProjectX builds Infinity, a workspace in the browser where every app (Blender, Unreal, Isaac Sim, VS Code and more) runs on its own dedicated GPU machine. Think cloud gaming, but for software.
    • Core platform — helped build the underlying code that still runs the product today.
    • GPU VNC backend — worked on the backend architecture behind it: GPU-backed VNC sessions where each app gets its own GPU in its own sandboxed environment, streamed from the server to your laptop, where you control it as if it were running locally.
    Show 1 moreShow less
    • Early-stage ownership — all-hands role in the company's earliest days, covering engineering and the business side of building a startup.

Featured project

MeldBuilt for the Meta AI Developer Hackathon

Slack and Jira in one app, with AI agents as teammates.

Teams decide things in chat and track them in tickets, and the work gets lost in between. In Meld, agents sit in your channels and keep the work up to date on their own: "I'll do it by Friday" becomes a ticket, "it's up for review" moves one, and a decision that gets reopened is linked back to where and why it was made.

Meld: Atlas logs a pricing decision in #product and updates ticket ORB-5, next to a mission tracking the paid-plans launch
~1 sto read every message (Muse Spark 1.2)
32 / 32agent eval cases passing
31 → 8 smission planning, tuned per call
Listens, then actsCommitments become tickets, progress said in chat moves them, small talk is ignored. Every change comes with Undo.
Remembers decisionsReopen a settled question and Atlas replies with who decided, when, why, and a link to the original thread.
Missions, not just ticketsGive it a goal: it plans the tickets, runs check-ins, re-plans when work slips and asks you only when a call is needed.
Bring your own agentAny MCP client (Claude Code, Codex, Cursor) joins as a workspace member with an inbox, channels and ticket access.

Agents only see what the person asking can see, so roles and permissions carry over to every action.

Next.js 16TypeScriptMeta Model APISupabaseReact QueryMCPSwiftUI (Mac)

Talks & travels

Papers, posters and talks that took me from Kolkata around the world. Scroll and the globe follows along.

India Kolkata Japan Tokyo5,100 km
  1. 01JapanTokyo, JapanDec 10, 2026

    PyTorch Day Japan

    Accepted · upcomingSpeaker

    “Five Bits on Metal: Adding Q5_K Quantization to ExecuTorch's MLX Backend”with Anika Tibrewala · PyTorch Ecosystem track

    A one-day PyTorch event in Tokyo, run by the Linux Foundation.

    ExecuTorch's MLX backend couldn't run Q5_K, one of the most common GGUF formats. We'll walk through adding it: the Q5_K super-block layout, a mat-vec kernel for decode and a tiled mat-mat kernel for prefill, and benchmarks of Q4_K, Q5_K and Q6_K on Apple silicon. Plus what it takes to get a kernel merged into a PyTorch runtime.

  2. 02KenyaNairobi, KenyaNov 20, 2026

    MCP Dev Summit Nairobi

    Accepted · upcomingSpeaker

    “Multi Round-Trip Requests – The New Safety Pattern in MCP Agents”with Anika Tibrewala · MCP Protocol in Depth track

    The Linux Foundation's summit for the Model Context Protocol, in Nairobi.

    Multi Round-Trip Requests are the new core pattern in the 2026-07-28 MCP spec for safe, auditable agent interactions: the server says it needs input, the client retries, and a request state ties the attempts together. We'll build an approval workflow on it in an MCP server and cover the pitfalls: state correlation, idempotency and timeouts.

  3. 03USASan Jose, USAOct 20, 2026

    PyTorch Conference North America

    Accepted · didn't attendPoster

    “Five Bits on Metal: Adding Q5_K Quantization to ExecuTorch's MLX Backend”with Shivay Lamba · Community Expo

    The PyTorch Foundation's flagship conference, in San Jose.

    Our poster on bringing Q5_K, a GGUF quantization format, to ExecuTorch's MLX backend so it runs on Macs was accepted for the Community Expo. I didn't make it to San Jose, but getting a poster into PyTorch Conference was a big one. The same work became my talk at PyTorch Day Japan.

  4. 04ChinaShanghai, ChinaSep 2026

    MCP Dev Summit Shanghai

    SpeakerScholarship
    Subhro at the podium with the title slide: How We Built Reliable Self Evolving Agent Skills
    The full room at AGNTCon + MCPCon China during Subhro's talk

    “How We Built Reliable and Self-Evolving Agent Skills”with Jayita Bhattacharyya · Agentic Engineering track

    AGNTCon + MCPCon China, the Linux Foundation's summit for the Model Context Protocol and agentic AI.

    Almost every layer of an agent stack is prompt-driven, and hand-tuning prompts against a non-deterministic model doesn't scale. We showed how agents can learn from their own mistakes: reflection-driven prompting that evolves behavior without RL or fine-tuning, with evaluations and feedback loops to measure which change actually made the system better.

  5. 05JapanHiroshima, JapanAug 2026

    PyCon JP 2026

    Scholarship

    Japan's largest international Python conference, held at the International Conference Center Hiroshima, inside Peace Memorial Park.

    My first PyCon JP: I attended on a scholarship, joining Japan's Python community in Hiroshima, a world away from Kolkata.

  6. 06USALong Beach, USAMay 2026

    PyCon US 2026

    PosterScholarship
    Subhro standing next to his poster at PyCon US 2026
    Subhro at the PyCon US 2026 entrance in Long Beach

    “Cost & Energy of Everyday LLM Workloads: A Visual Field Guide for Python Developers”with Shivay Lamba

    The world's largest Python conference, and my first PyCon.

    We profiled everyday LLM calls (chat completions, summarization, classification, RAG queries and batch embeddings) across model sizes, then turned the hidden numbers into visuals: latency waterfalls, cost per 1,000 tokens, and CPU vs GPU energy. Everything is plain pandas and matplotlib, so anyone can plug in their own endpoint and regenerate the charts.

    Off the poster floor: meeting the community, meeting Python's creator, and a weekend in Laguna Beach.

  7. 07USATacoma, USAJul 2024

    SciPy 2024

    Paper

    “Mamba Models a possible replacement for Transformers?”with Rounak Sen and Saikrishna Devendiran

    The Python in Science Conference, where the scientific Python community meets every summer.

    Our paper puts Mamba's state-space models side by side with Transformers: how the architectures differ, how they perform, and the mechanisms underneath. It's published in the Proceedings of the 23rd Python in Science Conference.

Also: merged PRs in PyTorch, vLLM and other major repos.

Open source

Recent merged and open pull requests, pulled live from GitHub for @JaynouOliver — this list grows on its own.

Papers, talks, conferences, and notes.