Four packages, fixed scope.
All packages ship as deployment binaries with documented configuration and a defined support contract. Pick the closest match and we'll align scope by email.
Delivery model
Implementations ship as deployment binaries (containerized or native) with an open configuration layer — schemas, API specs, runbooks, and parameter files are documented and editable. Core algorithms, models, and optimizations are not part of the deliverable. Every engagement includes a defined support contract (保守契約) for updates and fixes. Source code escrow available on request for procurement-sensitive buyers.
Foundation
Production-ready ingestion + calibration pipeline on the data you already collect.
A real-estate agency whose listing-feed updates from a dozen partners arrive in a dozen incompatible schemas — schema-normalized ingestion with calibration on a real day's traffic, queryable as one source.
Foundation builds the data pipeline you don't have time to build yourself: schema design, streaming or batch ingestion with proper backpressure, and a calibration sprint on real sample data. The output is a columnar storage layout that survives 10× growth, instrumented with per-stage timings and correctness checks wired into your observability stack. Hire Foundation when your current pipeline is taking too long, breaking under TB-scale, or producing data you don't trust. The 4–6 week engagement covers schema + ingestion + calibration + handoff. Two follow-up sessions included; longer support optional.
Discuss this project shape →Deliverables
- — Schema and columnar storage layout deployed to your environment
- — Streaming or batch ingestion with backpressure + retries
- — Calibration sprint on a representative sample dataset
- — Per-stage timings + correctness check suite, wired to your observability stack
Timeline
- Week 1 — Discovery + schema design
- Week 2–3 — Pipeline implementation + calibration
- Week 4–5 — Hardening + handoff
You provide
- — Sample data access
- — Calibration target
- — Reviewer time
I provide
- — Pipeline binaries + schema docs + runbooks
- — Methodology writeup
- — Two follow-up sessions
FAQ
- Can you work with our existing storage?
- Yes — most modern columnar formats, plain object storage, or your warehouse.
- Do you do operational handoff?
- Deployed binaries + runbooks + two follow-up sessions. Annual support contract structured separately.
Visualization & Query
TB-scale data made browsable: tile servers + NL query over your warehouse.
A property listing platform where customers fight a filter-cage UI (price + station + size + commute) to express fuzzy preferences like 'family-friendly, near a green line, 30 minutes from Shibuya, under ¥200K.' NL query over the listing data accepts the question as written.
Visualization & Query takes the data your Foundation produced and makes it browsable: a high-performance tile server backed by your columnar storage, a modern browser-based geospatial visualization, and a natural-language query layer wired to your existing warehouse. Hire this when your team is writing queries just to ask basic questions of geospatial data, or when your dashboards can't keep up with TB-scale joins. The 4–8 week engagement covers data inventory, UI shape, tile server, browser app, NL query layer, and a latency budget that's honest about the cost of joins. Includes the methodology writeup and follow-up sessions.
Discuss this project shape →Deliverables
- — High-performance vector tile server backed by your columnar storage
- — Modern browser-based geospatial visualization with configurable layers
- — Natural-language query layer wired to your existing warehouse
- — Latency budget + caching strategy
- — Curator (generative dashboards): a natural-language prompt assembles a validated dashboard (KPI / chart / table / map)
Timeline
- Week 1 — Data inventory + UI shape
- Week 2–4 — Tile server + browser app
- Week 5–8 — NL query layer + hardening
You provide
- — Read access to data
- — Domain reviewer
- — Auth/SSO requirements
I provide
- — Tile server + browser app + NL query binaries
- — Performance report
- — Two follow-up sessions
FAQ
- What zoom levels are supported?
- All of them. We pre-aggregate where it makes sense and stream raw at high zooms.
- Can the NL query layer use our schema?
- Yes — it maps your schema to safe, read-only queries against your warehouse.
Agentic Layer
Domain-specific agents and the orchestration substrate they run on.
An energy site-selection workflow where agents auto-investigate each candidate lot — 'is it grid-connectable, is the substation at capacity, what is EV charging demand within 2 km' — with citations to the source data instead of analyst hours per site.
Agentic Layer gives you domain-specific AI agents and the orchestration substrate they run on — not generic chatbots that hallucinate over your data. Hire this when off-the-shelf AI assistants are making confident-sounding claims about your domain that aren't backed by your actual data, or when you want to compose tool-using agents that operate against your real systems. The 6–10 week engagement delivers a tool server exposing your domain capabilities, an orchestration layer with structured I/O and tracing, an evaluation framework with golden + adversarial test sets, and cost + latency telemetry per run. No framework lock-in.
Discuss this project shape →Deliverables
- — Tool server exposing your domain capabilities to any agent runtime
- — Agent orchestration with typed tool I/O, retries, and traces
- — Eval harness for agent behaviors with golden test sets
- — Cost + latency telemetry per agent run
Timeline
- Week 1–2 — Tool inventory + interface design
- Week 3–6 — Orchestration + first agent
- Week 7–10 — Evals + telemetry + hardening
You provide
- — Domain expert time
- — Existing tools/APIs
- — Eval criteria
I provide
- — Tool server + orchestrator binaries + eval harness
- — Eval report
- — Two follow-up sessions
FAQ
- Do you use a specific agent framework?
- Standard tool-server protocol + your choice of runtime. I avoid framework lock-in.
- How do you measure agent quality?
- Golden set + adversarial set + cost/latency per run, all in the eval harness.
Microsimulation
Vehicle-level traffic simulation for scenario analysis, synthetic data, and what-if planning.
A property developer modeling 'what does opening this new shopping center do to neighborhood traffic' before construction commits the capital, with side-by-side metric comparison across alternative site plans.
Microsimulation builds vehicle-level traffic models for your geography using SUMO — the open-source microsimulator developed by the German Aerospace Center (DLR) and licensed under EPL-2.0. Hire this when you need to evaluate a network change before implementing it, when you need synthetic trajectory data for model training without real probe data, or when you need scenario-comparison metrics like travel time, throughput, and emissions. The 2–6 week engagement covers network build, demand calibration, scenario configuration, and synthetic output post-processing. Tested at city scale on Tokyo and Osaka networks.
Discuss this project shape →Deliverables
- — Road network calibrated to your geography (OSM-based)
- — Demand model fitted to your real counts, OD matrices, or sample probe data
- — Baseline + N alternative scenarios with side-by-side metric comparison
- — Synthetic trajectory output ready for downstream pipelines (map-matching, OD extraction, analytics)
Timeline
- Week 1 — Network build + data ingestion
- Week 2–3 — Demand calibration + baseline run
- Week 4–6 — Scenario configuration + comparison
You provide
- — Reference geography (city/region)
- — Calibration data (counts, OD, or probes)
- — Scenario specifications
I provide
- — Simulation binaries + calibrated configuration
- — Methodology writeup
- — Two follow-up sessions
FAQ
- Can you use real probe data for calibration?
- Yes — probe data feeds into the demand calibration step. The pipeline cleans and aggregates it before fitting.
- What scale can you handle?
- City-level: millions of vehicles per simulated day. Tested on Tokyo and Osaka networks.