Built withJabrod

Jabrod — 2026

Jabrod

Agentic AI platform

01

The studio's own platform: RAG pipelines, LLM agent systems and the infrastructure around them, built for teams shipping to production.

• /01Ingestion and parsing pipeline on GCS-backed storage
• /02Document chunking with per-knowledge-base identifiers
• /03Embedding generation into a vector store for fast retrieval
• /04Extensible agent orchestration for multi-step workflows

Juristo AI Labs — 2026

Juristo AI

Legal AI workspace

02

The AI half of Juristo's legal ecosystem — research, clause extraction and contract drafting, where retrieval accuracy is the product rather than a feature.

• /01Four-stage RAG pipeline: parsing, chunking, embeddings, retrieval
• /02Context-grounded legal responses rather than free-running generation
• /03LLM inference via OpenRouter behind a multi-agent routing layer
• /04Queries dispatched to the appropriate tools and workflows
• /05Prompt caching plus selective routing of lighter workflows

Juristo AI Labs — 2026

Juristo Lawyers

Legal marketplace

03

The marketplace half: clients search verified advocates for consultation, document review and contract help, and advocates earn from paid consultations.

• /01Production Next.js platform spanning APIs, auth and frontend delivery
• /02Background jobs carrying the work that cannot happen in a request
• /03Verified-lawyer discovery across practice areas and geography
• /04Paid consultations, document reviews and wallet tracking

Forward and backward attention kernels written from the papers, using online softmax so tiled attention stays numerically exact without ever materialising the N×N score matrix in HBM.

• /01FlashAttention-2 forward and backward kernels written in Triton
• /02Online softmax with running max/sum statistics keeps tiles exact
• /03Activation memory cut from O(N²) to O(N) by recomputing probability tiles
• /04Backward pass recomputes from a saved log-sum-exp vector
• /05Exactness validated against an fp32 oracle across fp16 and bf16
• /06Autotuned harness reporting latency, TFLOP/s and peak memory vs torch SDPA

A plan → verify → replan runtime that compiles a natural-language goal into a validated DAG of tool calls, and refuses to execute a plan it cannot check first.

• /01Rejects hallucinated tools, mistyped arguments and dependency cycles
• /02Validation happens before any side effect runs, not after
• /03Dependency-aware parallel execution with type-preserving dataflow
• /04Three-layer verification cascade: structural, then type, then LLM judge
• /05Bounded, evidence-driven replanning that reuses completed steps
• /06Eval harness separating engine metrics from planning quality

Industrial property across the NCR, plus the part nobody else covers — the subsidy, finance and approvals work it takes to actually get a factory running on the land.

• /01Listings for industrial plots, sheds and warehouses to buy or rent
• /02Subsidy and incentive finder with free eligibility checks
• /03Project cost estimator for pre-engineered and RCC construction
• /04Approvals roadmap covering allotment, filings and board consents

IBSEA — 2026

IBSEA Professional OS

Portal and marketplace

07

Verified member profiles, professional discovery, consultation intake and events — plus consent-led AI representatives that speak for a member without ever cloning them.

• /01Role-based hub for Institute, Talent and Project accounts
• /02Consultation intake, community collaboration and event RSVPs
• /03Talking representative with explicit per-profile voice consent
• /04Administrator governance with no silent fallback to sample data