Speaker
Description
FireForm is a UN-OICT official Digital Public Good , that lets first responders generate every agency's required PDF reports from a single voice memo — no manual form-filling, no duplicate data entry across departments. I'm building this as a GSoC 2026 contributor, selected from over 200 applicants for one of two spots.
The pipeline has three core pieces: OpenAI Whisper handles local voice transcription with preprocessing and abbreviation normalization (fire department jargon doesn't transcribe cleanly out of the box). Pydantic validation with retry logic, layered on Ollama and Mistral 7B, catches and corrects extraction errors instead of letting them fail silently. A YAML-driven multi-agency PDF system lets a new fire department onboard by writing a config file, not code.
This talk walks through how each piece works, the specific failure modes you hit when LLM output feeds directly into structured, agency-specific documents, and what it's actually like contributing to an active open source project as a GSoC contributor — how to get useful context fast in an unfamiliar codebase, when to ask maintainers versus dig in yourself, and what it takes to get a contribution merged into production software used by real fire departments.
No ML background required. This is a talk about engineering discipline around LLMs in production, not model architecture.
What you'll leave with:
- A concrete look at how voice-to-structured-document pipelines are built
- Specific failure patterns when LLM output needs to be reliable, not just plausible
- A realistic picture of what GSoC contribution work looks like, for students considering applying
Any other info we should know?
This talk is technical in nature — it covers a real production pipeline (Whisper for transcription, Pydantic validation with retry logic on Ollama/Mistral 7B, YAML-driven PDF generation) and is aimed at people comfortable with basic programming concepts. No ML background is required; the focus is on engineering decisions around making LLM output reliable, not on model internals or training.
What the audience gets:
- A concrete example of a voice-to-document AI pipeline built for a real client (UN-OICT, used by US fire departments) — not a toy demo
- Specific lessons on validating and retrying LLM output so failures don't pass through silently
- A practical view of what GSoC contribution work actually involves, useful for students who are considering applying to GSoC or similar open source programs
Session author's bio
Abhishek Kumar is a Computer Science graduate from Dr. A.P.J. Abdul Kalam Technical University, currently a GSoC 2026 contributor at the United Nations Office of Information and Communications Technology (UN-OICT), building a production AI pipeline for FireForm - selected from over 200 applicants for one of two spots. He previously worked as an LFX Mentee with the Cloud Native Computing Foundation (CNCF) on the KubeEdge Ianvs benchmarking toolkit, and as a Software Development Engineer Intern at Akatsuki in Japan, building a production AI-powered translation platform. He's an ICPC Asia West Regionalist
| In Person Attendance | In-person |
|---|---|
| Please confirm that there are included headshots of all speakers in their profiles | Yes |
| Agree to Privacy Policy and Notice | I agree |
| Social Media | linkedin.com/in/abhishekkumarji |
| Level of Difficulty | Intermediate |