Projects
25 selected pieces of work, newest first. Products, agent systems and machine learning competitions. 4 of them won.
HALFSPREAD
AI Trading Agents Hackathon, lablab.ai and Alpaca
Every options agent knows how to enter a trade. Almost none price what it will cost to get out. HALFSPREAD measures execution cost before every order, ranks structures by what survives it, and settles positions rather than paying the spread twice. That matters more than the signal: crossing costs 0.52% on NVDA and 33% on a QQQ contract with 26 minutes left. Every published figure re-derives from an append-only journal with no key and no network, and the build fails if one stops reproducing.
Cassandra
All Things Agentic Hackathon
Spreadsheets are the largest untested codebase on earth, and 94% of them contain an error. Cassandra is continuous integration for them. Drop a workbook in a bucket and an agent fleet wakes from zero, builds a formula dependency graph, writes the fix, and proves it by recalculating the file. Any correction that moves the wrong cell is rejected with its reason. On the demo model it caught a workbook claiming 6.2M in operating income for a company that had lost 1.7M, and repaired five defects in 114 seconds unattended.
Cascade
CockroachDB and AWS Hackathon, Build with Agentic Memory
The problem with an automated runbook is not finding it, it is knowing whether it is still true. Nothing about a stale procedure looks stale: it matches the incident, executes cleanly, and does the wrong thing with confidence. Cascade pins every learned procedure to the exact version of each policy rule behind it, so changing one rule invalidates all its dependents in four writes, whether that is one procedure or a hundred thousand. Staleness is a join, not a column.
Gotcha!
Prometheus July AI Challenge
An AI that lies on purpose, so students learn to catch it. Every challenge hides exactly one deliberate mistake. You find it, explain why it is wrong, and say what it should have been, and a second model on a different provider marks your reasoning. Separating generation from grading is what makes the score worth anything: the system never marks its own homework. The goal is not to get answers out of a model, it is to stop taking them on trust.
AwardCADENCE
Winner, Best Overall Project, ML Empowerment Build Challenge 2.0
Parkinson's screening from a 30 second voice sample, with an explanation attached to every score. Three tasks feed the model: a reading passage, a diadochokinetic rate test, and sustained vowel phonation for jitter, shimmer and harmonics-to-noise ratio. The classifier is validated across three independent language corpora using cross-database testing and domain-adversarial training, which is what separates a real result from the microphone bias that inflates most published voice screening work. Ships as an installable multilingual PWA.
Olikobochon
IUTCS Datathon 2.0, Institute of Policy Dynamics
A model can name the author of Bidrohi correctly in English and hand the poem to the wrong poet in Bengali, in prose that sounds completely fluent. This competition asks you to catch that. Given a Bengali prompt and a candidate answer, decide whether the answer is faithful or invented. My pipeline is a decision ladder: match against the public corpora the benchmark was built from, then fuzzy-match an 89k-question bank, then idiom lexicons, and only send the irreducible residue to an open-weight judge with Bengali Wikipedia retrieval behind it. Scored 0.922 F1 on the hallucinated class, third on the public leaderboard.
DeepSIFT
Find Evil, SANS DFIR Challenge
Forensic evidence is worthless if the analyst cannot show where a conclusion came from, which rules out a model that guesses at CLI output. DeepSIFT parses 148 real SIFT tools into structured JSON before the model sees a byte, tags findings against MITRE ATT&CK, scores confidence on four axes, verifies every claim against the evidence that produced it, and writes a signable chain of custody. Scored 4/4 against published ground truth on two cases with zero hallucinations and 100% claim grounding.
Zero
Waste gets reported and never collected because nobody is accountable for the gap. Zero closes it with money. A citizen reports with a photo, the model classifies the waste and the location snaps to a mapped zone by point-in-polygon check, a cleaner claims the job from a marketplace, and payment releases only after before-and-after image comparison plus a citizen review. Four roles, zone mapping, green points, leaderboards and withdrawals.
Appoint
Gemini Live Agent Challenge
Healthcare portals assume you already know which specialist you need. Appoint does not. Describe a symptom out loud in English or Bangla, or show it a photograph of a rash or an injury, and it asks its own follow-up questions, flags emergencies against an urgency model, reasons from symptoms to the right specialist with an explicit trade-off between distance, rating and facilities, then completes the booking. Built for the people portals leave behind.
AwardSolarSnap
Grand Prize Winner, FLIR App Challenge 2025 to 2026
A thermal image of a failing solar panel is useless if nobody can say which panel it was. SolarSnap pairs an Android app with a FLIR ACE camera and binds every reading to a specific physical asset, syncing to a Flask backend that turns captures into fault reports and maintenance history. Removing the manual reconciliation step is what makes large-array inspection viable, and it recovers yield without any additional hardware on site.
DL Sprint 4.0
DL Sprint 4.0, BUET CSE Fest 2026
Bangla speech recognition on long-form audio: lectures, interviews and real conversations rather than clean single-speaker clips. The difficulty is everything around the speech. Recordings run for hours, speakers overlap, music bleeds in, and silence stretches long enough that a decoder will happily hallucinate through it. Scored on word error rate weighted by sentence length, so a long sentence transcribed badly costs what it should. Three weeks of training runs against a hidden test set.
Resonate
Build your Flutter Butler with Serverpod
Mood trackers ask how you feel. Resonate listens to how you say it, combining acoustic and semantic analysis across English and Bengali so the signal does not depend on self-report. Four privacy levels govern what ever leaves the device, and the tracked signal feeds journaling, meditation and long-run mental health trends rather than a single daily number.
DoNotMiss
Codegeist 2025, Atlassian Williams Racing Edition
Work gets assigned in passing, in an email or a chat message, and then lives or dies on somebody remembering it. DoNotMiss catches it where it is said and turns it into a Jira issue without leaving the page. Detection is automatic, creation is not: every issue waits for approval and carries its source link and timestamp, so the context survives into the backlog instead of being reconstructed from memory a week later.
AwardDoclyst
Winner, Best ERNIE Multimodal Application, ERNIE AI Developer Challenge by Baidu
A patient holding a lab report has every number and no idea which one matters. Doclyst reads the report, extracts the values with OCR, and returns a plain-language explanation with safety alerts in under 30 seconds, so the thing that needs attention is the thing that stands out. Built as a safety interpreter, not a diagnosis tool.
AwardCoven
Winner, Honourable Mention, LMA EDGE Hackathon, showcased at the London finale
Loan covenants are buried in documents nobody rereads until one is breached. Coven turns the documentation into a visual timeline with live covenant status, risk signals and explanations a non-specialist can act on. Built for the Loan Market Association's global hackathon on how loans are originated, documented, traded and managed, and presented at the LMA EDGE finale in London.
Chimera Protocol
Kiroween
Switch model providers and your context resets, which quietly locks you into whichever one you started with. Chimera Protocol is a shared memory layer that survives the switch: conversations, documents and imported pages are vectorised once and injected into whichever model answers next, across OpenAI, Anthropic, Google and DeepSeek. Team workspaces, and a developer console for inspecting what the router actually did.
GenForge
FIBO Hackathon
Generated game art usually arrives as a picture, which is not the same as an asset. GenForge turns a text prompt into production-ready content through two engines, one for 2D sprite sheet animations and one for 3D models, both emitting formats that drop straight into Unity or Godot with no manual cleanup in between.
CareForAll
A healthcare and philanthropy platform built as seven services that genuinely deploy on their own: authentication, campaigns, pledges, payments, totals aggregation and chat behind an API gateway, with a Next.js frontend over all of them. Built as a polyrepo with real service boundaries and its own Prometheus and Grafana stack, specifically to learn where that architecture pays for itself and where it does not.
Emotuna
Autocomplete writes like everyone. Emotuna writes like you. It learns one person's slang, tone and emoji habits and drafts replies in that voice across Telegram, Discord, WhatsApp and more. Messages are classified for emotion, sentiment and toxicity, that feedback trains a per-user Direct Preference Optimization model, and each user's data stays strictly isolated. Routine replies go out automatically; anything that matters waits for approval.
Shomonnoy
The water utility, the gas utility and the roads department cut the same street three times a year because none of them can see the others' plans. Shomonnoy puts every scheduled work on one shared map and surfaces the clashes while they are still on paper, which is the only point at which fixing them is cheap.
QuranWhispers
Keyword search is the wrong tool when you know how you feel but not what you are looking for. QuranWhispers searches verses by emotion and theme instead, plays and requests recitations, generates a daily dua and shareable verse posters, and runs an interactive forum driven by smart commands with a full admin panel behind it. I designed and built the JavaFX frontend from scratch.
Srabon
SrotDev's first project, and still the one I am fondest of. Srabon builds a personalised science course for a secondary schooler from a short survey, then teaches it through storytelling and quizzes rather than exposition. Multilingual and accessibility support are in the design from the start, so a language barrier or an impairment does not decide who gets to learn.
KugelBlitz: Pace in the Classroom
NASA Space Apps Challenge 2024
A classroom astronomy resource that pairs a multi-level game with the study material so the exploration and the syllabus reinforce each other instead of competing for attention. I designed the interface during the NASA Space Apps Challenge 2024.
CodeInception
A customisable bootcamp that more than 50,000 young programmers used as a starting point, run with a team of domain experts who led the sessions. Organising it taught me more about explaining things clearly than any amount of writing code alone ever did.
Hospital Management System
A cross-platform desktop application covering the administrative, medical and billing workflows of a mid-range hospital, built with Kivy over a PostgreSQL schema. One of the first things I wrote that had to hold real state correctly rather than merely run.
