A single LLM step dominates a recurring cost line.
LLM & Small Language Model Engineering
Engineer the right language model before funding the build.
Our focused entry engagement—the Custom SLM Decision & Scaling Study—evaluates one production LLM workload across a realistic candidate ladder: optimization and routing, adaptation and fine-tuning, distillation, continued pretraining, and scaled scratch experiments.
Strong fit signals
- One stable, repeated workflow
- Representative examples or known failures
- Meaningful spend or a hard deployment constraint
- A decision owner who can approve trade-offs
Capability and entry offer
LLM & Small Language Model Engineering is a specialist capability within Mankash AI’s broader enterprise agent architecture practice. The focused commercial entry point is the Custom SLM Decision & Scaling Study below.
When to engage
Symptoms that justify a measured model decision.
The study is intentionally narrow. It is designed for a production workflow where quality and operating constraints can be made explicit.
Hosted APIs conflict with private, VPC, on-prem, or edge requirements.
Latency, memory, or throughput is blocking product behavior.
Provider changes or availability create material business risk.
Your team is debating fine-tuning, distillation, or pretraining without a common evaluation.
A narrow evaluator, classifier, router, or guardrail runs at high volume.
Study stages & artifacts
A fixed-scope path from workload definition to investment decision.
Typically 3–4 weeks after data and access are ready. A paid, fixed-scope engagement. Final scope follows a free technical consultation.
Define acceptance
Map the workflow, incumbent, failure modes, and constraints.
- Acceptance gates
- Readiness assessment
- Baseline plan
Build the harness
Create a reproducible workload-specific test set and measurement path.
- Evaluation harness
- Metric definitions
- Failure taxonomy
Test candidates
Move up the candidate ladder only when cheaper options miss the gates.
- Candidate runs
- Error analysis
- Scaling curves where relevant
Decide & scope
Compare quality, cost, latency, ownership, and lifecycle obligations.
- Pareto frontier
- 12-month TCO
- Go/no-go implementation plan
Candidate ladder
The next experiment must earn its cost.
- 01 Measure Establish the incumbent baseline and acceptance gates.
- 02 Optimize / reroute Improve prompts, context, caching, and model selection.
- 03 Adapt Test the smallest useful open-model adaptation.
- 04 Distill Transfer task behavior into a smaller candidate.
- 05 Continue pretraining Add domain learning only when the evidence calls for it.
- 06 Pretrain from scratch Run scaled experiments before any full training decision.
We stop at the least expensive approach that clears the agreed acceptance gates.
Required inputs
Enough evidence to make the result operationally meaningful.
A useful study normally requires one clearly defined workflow, representative inputs and expected outputs, known failure examples, current usage/cost estimates, and a decision owner who can approve trade-offs.
Raw customer data is not required before contracting and establishing an approved transfer process.
Do not send raw examples or confidential datasets through the consultation request form.
Acceptance gates
Customer-specific numbers, agreed before conclusions.
This template illustrates the comparison structure. It contains no invented baseline, target, or candidate result.
| Dimension | Baseline | Required gate | Candidate result |
|---|---|---|---|
| Task quality | Measured in discovery | Agreed with customer | Measured in study |
| Critical-error rate | Measured in discovery | Agreed maximum | Measured in study |
| P95 latency | Measured in discovery | Agreed maximum | Measured in study |
| Throughput | Measured in discovery | Agreed minimum | Measured in study |
| Cost per 1,000 tasks | Measured in discovery | Target range | Measured in study |
Commercial path
Separate the decision from the implementation commitment.
- Free 30-minute technical consultationNo charge
- Data/readiness checklistMutual NDA if needed
- Fixed-scope decision studyPaid engagement
- Results reviewExplicit go / no-go
- Separate implementation SOWOnly if justified
- Optional lifecycle supportOptimization and monitoring
Implementation is scoped separately after the study. Full 1–4B pretraining projects are not publicly priced before scope and compute are validated.
Proof & limitations
Relevant experience without pretending the next scale is already solved.
FAQ
Plain answers before the consultation.
Commercial, security, deployment, and ownership details are finalized in an executed agreement—not inferred from marketing copy.
Why not just fine-tune an open model?
Fine-tuning may be the right answer, and it is part of the candidate ladder. The study first establishes a workload-specific baseline and tests whether prompt/context changes, routing, or adaptation can clear the acceptance gates before considering more expensive approaches.
Will you always recommend a custom model?
No. We recommend the least expensive, lowest-risk approach that meets the agreed quality, cost, latency, and deployment gates—even when that means keeping the current system.
Can you guarantee savings?
No. Savings depend on the workload, volume, quality threshold, infrastructure, and model-lifecycle costs. The study is designed to measure those trade-offs before a larger commitment.
What data do you need before the consultation?
No raw customer data is required for the free consultation. A useful conversation needs a clearly described workflow, its constraints, approximate usage or cost, and an understanding of whether representative examples exist.
Can work happen in our environment or on premises?
It may be possible, subject to workload, security, access, and deployment review. The delivery location and operating controls are agreed in the statement of work; the website does not make a blanket security or residency commitment.
Who owns the code and checkpoints?
Ownership and permitted use of code, configurations, checkpoints, and other artifacts are defined explicitly in each executed contract. We do not impose a public default that could conflict with a customer’s requirements.
Have you trained a 1–4B model before?
The team has completed approximately 50 language-model training runs below 500M parameters. Mankash AI has not yet delivered a 1–4B production pretraining program, so larger-model work begins with smaller scaling runs and explicit technical and commercial gates.
What happens if no candidate beats the current system?
We document the evidence, failure modes, and economics, then recommend keeping or improving the current system. Avoiding an unjustified build is a successful study outcome.
Do you provide compute?
The in-house lab supports rapid experiments. Larger compute is sourced only after an approved scaling plan defines the hardware, budget, security, and delivery requirements.
Free consultation request
Tell us about one workflow.
Share enough operational context for us to assess whether a free consultation would be useful. Do not send examples, datasets, customer records, or credentials yet.
What happens next
We review the workload and reply within two business days. If there is a potential fit, we arrange a free 30-minute technical consultation.
Do not include customer records, credentials, proprietary datasets, or other confidential information.
Please review the highlighted fields.
Thank you.
We will review the workload and reply within two business days. Please do not send data until we agree on a secure process.