Location: Brazil (UTC-3)
Remote: Yes. Happy to overlap substantially with US/Canada, European, Australian or New Zealand working hours.
Willing to relocate: No
Technologies: Python | LangGraph | LangChain | Agentic AI | Graphiti/Neo4j | RAG/graphRAG | PyTorch/HuggingFace | n8n | AWS | Terraform | Docker | FastAPI | PostgreSQL | React/Next.js | MCP | Claude Code and Codex (with engineered harnesses, solid human judgement and experience arguing with them)
Résumé/CV:
General tech: https://drive.google.com/file/d/1igF38vIVHMa-bghFMPC3MfimBtV...
Health/clinical/biomed domain-specific: https://drive.google.com/file/d/1FUUemGkPPDOy0YKFG1FHQpCfzme...
Email: mprodhi@gmail.com
GitHub: https://www.github.com/prodm93
Whitespace demo (frontend accessibility still needs work): https://drive.google.com/drive/folders/18UvWT5NriOlVy54Yz5KC...
AI engineer, agentic systems builder and former biomedical researcher. I have 5+ years of development experience and 4+ years building AI systems professionally across startups and larger corporate clients, with work spanning biomedical/clinical AI, finance, content systems, automation and general-purpose LLM applications (including regulation- and compliance-heavy domain work).
The common thread is probably best described as “grounded mad scientist” energy. I like getting dropped into problems where there is no roadmap, the inputs are horrible, the requirements are fuzzy and the straightforward-looking solution stops being straightforward approximately fifteen minutes after you touch the real data. I tend to do my best work somewhere around that point.
A few examples:
For a long-term startup client, I effectively became the technical founding person with no repo, no technical guidance and no roadmap in sight. Among other things, I built their AI-driven content and automation ecosystem from scratch. When generic AI output became the problem, I came up with an in-place DPO approach that mined their actual messy Notion editing history, including highlights, strikethroughs and collaborative edits, to generate positive/negative preference pairs for few-shot prompting. No labellers and no fine-tuning. A/B testing, which I also designed for non-technical founders, showed more than 30% improvement in output quality.
For Danaher Corporation via NILG.ai, I built a pharmaceutical pipeline extracting clinical endpoints from FDA drug labels and ClinicalTrials.gov records, then classifying them into higher-level outcome categories. The extraction itself was the easy bit. The interesting work started when the FDA API confidently returned the wrong drug label, source datasets had gaps and seemingly simple client terminology turned out to be inconsistent enough to silently poison downstream analysis. A lot of my work is exactly this: translating fuzzy human requirements into technical systems that are not merely impressive-looking, but actually trustworthy.
For Berkshire Partners, I built a private long-context retrieval and summarisation system over confidential 120-page interview transcripts using locally-run open-source models. This was early 2024, under severe context-window constraints, so I adapted LangChain's refine chain across multiple retrieval rounds to preserve information while generating higher-order outputs such as market forecasts and investment opportunities. Roadmap? Nowhere to be seen.
I also recently had a feature PR merged into Backblaze's open-source genblaze SDK. The docs told users in several places to hash fetched assets themselves for provenance verification, but no shipped tool actually did it. I added opt-in byte-level verification through the existing SSRF-hardened transfer path with DNS pinning, presigned URL credential redaction and per-asset failure isolation. What I enjoyed most was dropping into an unfamiliar codebase, figuring out why earlier architectural decisions had been made and building around those constraints instead of bulldozing through them.
I started my AI engineering career at an AI + NLP research startup called Sagewrite (before the ChatGPT hype took off), where I built substantial chunks of their production backend. I worked on scientific PDF parsing and text-processing pipelines, prepared model-training datasets and built scientific text-generation systems using models including GPT-2. So yes, I’ve been trying to get useful things out of AI slop since GPT-2, which had approximately the literacy of a toddler–work in this field is not a fad pivot for me!
As an erstwhile scientist, I also come with extensive biomedical research experience. My wet-lab background is in immunology, immunotherapy and cancer biology, including doctoral work at Amsterdam UMC and MSc research at CIC BioGUNE that contributed to a Nature Communications paper on Siglec-15. Since moving into AI, I have worked on pharmaceutical and clinical-trial pipelines, PubMed/citation analysis, therapeutic chatbot systems, RAG over biomedical literature and antiviral drug-discovery ML. I am very interested in biomedical/healthtech AI, but absolutely not limited to it.
Outside client work, I am currently building Whitespace, a React/Next.js app that maps a user's professional expertise against the patent landscape to surface unmet needs and generate validated R&D ideas. It uses agentic graphRAG, multi-agent workflows, an adversarial multi-LLM council and human-in-the-loop review. The backend is being built as a proper production system rather than a demo held together with hope: Terraform-defined AWS infrastructure, LangGraph workflows, async orchestration, observability and a lot of thought around security. Demo videos/screencaps linked above.
The other useful thing to know about me is that I learn obscenely fast. My brain usually has about 2048 tabs open, most involving some Google rabbit hole, documentation page or StackOverflow thread explaining why the thing that "should obviously work" does not. I use AI heavily in my coding workflow too, but I am perfectly happy arguing with it until I understand why a design is sound rather than accepting whatever compiles.
Logistics:
I work through Confluente Ltda, my registered Brazilian PJ entity. For overseas companies this can be structured as a straightforward B2B vendor relationship rather than foreign payroll or visa sponsorship. Day to day, I show up, do the work and remain accountable like any other member of the team. Administratively, there is much less cross-border employment machinery for you to deal with.
I have worked remotely with clients across multiple continents and time zones for years, including long-running engagements, so async work and deliberate communication are very normal for me.
Open to full-time, part-time or long-term contract arrangements. Special place in my heart for founding eng roles. Interested in AI/ML engineering, agentic systems, AI-native software development, automation, applied AI and biomedical/healthtech/scientific AI.
English is my first language (I'm an anglophone transplant to Brazil).