Nexera
Built Nexera — an autonomous AI research workspace that plans queries, gathers evidence across the web, and ships citation-backed reports with live streaming progress.
Nexera is an autonomous AI research workspace that classifies queries, plans sub-questions, executes searches, fetches and chunks web content, reranks evidence, and synthesizes citation-backed reports — all while streaming progress live. It supports quick, standard, and deep research modes, thread-aware follow-ups, URL/file ingestion, and multimodal capabilities (speech transcription, TTS, image analysis).
Project At A Glance
Timeline
2026
Industry
AI · Research · Agentic Systems
Contribution
Recursive-RAG pipeline, thread-aware memory, SSE streaming, multimodal ingestion, and full-stack delivery
Collaboration
Project Visual
Nexera
Summary
What this project demanded.
Nexera is an autonomous AI research workspace that classifies queries, plans sub-questions, executes searches, fetches and chunks web content, reranks evidence, and synthesizes citation-backed reports — all while streaming progress live. It supports quick, standard, and deep research modes, thread-aware follow-ups, URL/file ingestion, and multimodal capabilities (speech transcription, TTS, image analysis).
Role
Full Stack Developer
Year
2026
Problem Space
Autonomous research products break when retrieval drifts, threads forget context, or the user can't see what the system is doing. Nexera had to keep recursive-RAG accurate, threads coherent across follow-ups, and progress legible in real time.
I built a recursive-RAG pipeline with separated context layers — recent chat, thread summaries, long-term memory, and trusted sources — plus SSE-driven live event streaming, markdown reports with citations and PDF export, and BYO API keys with model discovery so users can plug in their own providers.
Capability 01
Next.js
Capability 02
TypeScript
Capability 03
FastAPI
Capability 04
Python
Capability 05
MongoDB
Capability 06
SSE
Capability 07
RAG
Capability 08
Multi-Agent
Agentic Research
Nexera had to feel like a research partner, not a chatbot.
Autonomous research products only work when the retrieval, memory, and progress UI are designed together. My job was to make the pipeline transparent enough that users could trust the report — and the citations behind it.
01
Pipeline & Memory
- Recursive-RAG: query classification, sub-question planning, search, chunking, rerank, synthesis
- Separated context layers: recent chat, thread summaries, long-term memory, trusted sources
- Quick, standard, and deep research modes for different depth/latency trade-offs
02
Workspace & Inputs
- Live event streaming over SSE with markdown reports and PDF export
- URL and file source ingestion alongside web search results
- Multimodal: speech transcription, text-to-speech, and image analysis with BYO API keys
Signal
3 Modes
Quick, standard, and deep research depths over the same recursive-RAG pipeline
Signal
4-Layer Memory
Recent chat, thread summaries, long-term memory, and trusted sources
Signal
SSE Live
Real-time progress streaming with citation-backed markdown and PDF export
Insight
Autonomous research products live or die on retrieval and trust. Citations, memory, and visible progress are the product — not the chrome around it.
Core Interaction Shifts
The product decisions that changed how the experience felt.
Shift 01
Recursive-RAG Research Pipeline
Classifies queries, plans sub-questions, executes web searches, chunks and reranks content, then synthesizes citation-backed reports across quick, standard, and deep modes.
Shift 02
Thread-Aware Memory Layers
Separated context across recent chat, thread summaries, long-term memory, and trusted sources so follow-up research stays coherent without leaking state.
Shift 03
Live Streaming + Multimodal
SSE-powered live event streaming with markdown reports, PDF export, URL/file ingestion, speech transcription, TTS, and image analysis — all wired into a single workspace.
Influence & Validation
What changed because of the work.
Nexera shipped as a working autonomous research workspace with the retrieval, memory, and streaming foundations needed to feel like a real research partner.
Recursive-RAG pipeline with reranking and citation-backed synthesis across three research depths
Thread-aware memory across recent chat, thread summaries, long-term storage, and trusted sources
SSE live progress, PDF export, multimodal inputs, and BYO API keys with automatic model discovery