Forecasts with honest ranges
Agni Predict will turn your time-series data into p10 / p50 / p90 forecasts. We're building it on our own forecasting foundation model, to run in the cloud or fully inside your infrastructure. Try the statistical studio below on your own data today.
In development · Planned: UK / EU residency · Cloud or on-prem · No data egress in private mode
Forecast your own data in Agni Predict Studio
Drop in a CSV or JSON time series, or try a synthetic sample. The studio backtests a family of statistical models on your history, picks the one that was most accurate, and gives p10 / p50 / p90 forecasts you can export. Nothing is uploaded.
Statistical forecast computed in your browser: exponential smoothing and naive models, chosen by rolling-origin backtest. This is not the Agni Predict foundation model. Bands come from this series' backtest errors; on short, volatile or trending histories they tend to be too narrow.
Start in the cloud, move to your own infra
At launch: the same model and API across cloud, private deploy and your own fine-tuned checkpoint, so you can switch without rewriting a line of integration code.
Cloud Forecast API
Self-serve. Pay by the forecast.
- POST a CSV or JSON of historical series
- Receive point + p10 / p50 / p90 ranges
- Optional covariates (promos, weather, holidays)
- p95 latency under 3s for ≤100 series
Private Deploy
Docker image. Your VPC. Zero egress.
- Weights baked in, runs fully offline
- Helm chart for Kubernetes
- Per-node annual licence with updates
- Ideal for regulated and on-prem workloads
Domain-Tuned
Your data. Your private checkpoint.
- Upload 6–24 months of historical data
- Overnight LoRA fine-tune (1–4 GPU-hours)
- Private checkpoint, served from Cloud or Private
- Re-tune quarterly included
The boring stuff done properly
Ranges, not false precision
Every forecast comes with p10 / p50 / p90 quantiles. Your team sees the uncertainty, not just a single line pretending to be the truth.
Zero-shot, then domain-tuned
We're building a pretrained forecasting foundation model for zero-shot forecasts, with fine-tuning on your own data for a private checkpoint. Early-access customers will measure it against their current baseline before committing.
Your data stays where you put it
Cloud mode is planned for eu-west with no third-party telemetry, and private mode will run entirely inside your VPC. The studio on this page already runs entirely in your browser.
Backtest before you commit
The studio above already backtests every forecast on your own history and shows the accuracy it achieved. Early-access customers will get the same test against their current baseline before paying for a seat.
Built for decisions, not dashboards
Retail demand
Forecast SKU × location × week with promo and holiday covariates. Replace brittle Prophet pipelines with p10 / p50 / p90 ranges your replenishment team can actually trust.
SaaS & fintech KPIs
MRR, churn cohorts, cash runway, treasury forecasts. Honest p10–p90 bands replace overconfident single-point projections in your board pack.
Energy & utilities load
Half-hourly load forecasting with weather covariates. Self-hostable for grid operators and regulated reporting workloads.
Field ops & logistics
Call volume, ticket volume, delivery ETAs. Plan staffing and fleet capacity from the same forecasts your finance team will use for cash planning.
Annual-first. No hidden compute charges.
Planned launch pricing. Waitlist members get a 14-day free trial on the Starter feature set, with no card up front. Annual billing gets the prices below; monthly will be available at +20%.
Starter
Single team, up to 50 series
- 25,000 forecast credits / month
- Cloud Forecast API
- p10 / p50 / p90 ranges
- Email support, 1 business day
Growth
Multi-team, up to 2,000 series
- 250,000 forecast credits / month
- Covariates (promos, weather, holidays)
- Async batch jobs
- Shared Slack support, 4 hour SLA
- One backtest workshop with our team
Scale
Vertical specialist, domain-tuned
- 1,000,000 forecast credits / month
- One LoRA fine-tune per quarter
- Private checkpoint, served via your tenant
- 99.9% uptime SLA, named CSM
Private Deploy
Regulated and on-prem workloads
- Docker image + Helm chart
- Per-node annual licence
- Quarterly model updates included
- Named customer success manager
A forecast credit = one series, one forecast call, up to horizon 90 and context 1024. Larger horizons and contexts consume proportionally more credits. Full credit math published in the docs once Agni Predict goes GA.
Clear roadmap. Honest posture.
Our own model, in development
We're building Agni Predict on our own pretrained forecasting foundation model. Any private checkpoint fine-tuned on your data will be yours outright.
UK & EU data residency
Cloud mode is planned to run in eu-west by default. Private Deploy will run entirely inside your own infrastructure for US, APAC, or regulated workloads where data residency is mandatory.
Aligned with UK GDPR
Article 13 notices, DPA on request, SCCs for any sub-processor. We do not pursue SOC 2 or ISO certifications at launch and never claim them — we phrase posture as 'aligned with', not 'certified'.
Honest about the limits
The model is univariate at launch. Very short or intermittent series get flagged, not silently forecast, as the studio above already does. Covariate pipelines will be audited against known model limits before any release.
Join the Agni Predict waitlist
Tell us about your forecasting work. Design partners get six months of Growth tier in exchange for a case study.