Bringing Trustworthy AI On-Premises: How AlphaNimble and ASM Technologies Put Memuron to Work on Technical Documentation

Engineering organizations sit on some of the most valuable knowledge anywhere: design specifications, test procedures, part catalogs, manufacturing notes, and years of accumulated project documentation. Making that knowledge searchable with AI is an obvious win. Doing it without sending sensitive technical material outside the building is the hard part.

That was the challenge ASM Technologies Limited brought to AlphaNimble. The result was a fully local, on-premises deployment of Memuron that ASM tested independently in its own environment, reaching an average Overall Trust Score of 84.96 and finding the required evidence for every single evaluated query.

Here is how the engagement came together, what we built, and where it goes next.

About ASM Technologies

Founded in 1992 and headquartered in Bengaluru, ASM Technologies is a publicly traded global engineering services provider specializing in Engineering Research & Development (ER&D) and Design-Led Manufacturing (DLM). The company supports the full product development lifecycle, from design, prototyping, and testing of complex mechanical, electronic, and embedded systems to Industry 4.0 automation, IoT, robotics, and sustenance engineering for mature products.

ASM serves demanding verticals including semiconductor and electronics, automotive and EV engineering, renewable energy and solar, and enterprise software, with delivery centers and offices across India, the USA, the UK, Singapore, Japan, Canada, and Mexico.

For a company like this, technical documentation is not a side asset. It is the business.

The Challenge: Document Intelligence That Stays Local

ASM needed a document intelligence stack capable of retrieval and generation over technical material, with one non-negotiable requirement: storage, retrieval, embeddings, and model execution all had to remain under ASM's control.

That rules out the usual shortcut of piping documents into a hosted AI service. It also raises the bar on retrieval quality. Engineering questions rarely look like casual web searches. A single query might combine natural language ("what is the torque procedure for...") with exact part numbers, filenames, revision codes, or domain-specific terminology. A system that only understands semantic similarity will miss the identifiers; a system that only matches keywords will miss the meaning.

The Solution: A Self-Contained Memuron Stack

AlphaNimble adapted Memuron for ASM's operating model and delivered it as a single, reproducible Docker Compose workflow. Once the required images and models are downloaded, routine operation stays entirely inside the client's environment.

The package brings together:

• Memuron as the retrieval and knowledge layer
• PostgreSQL with pgvector for retrieval metadata and vector search
• MinIO for document object storage
• Ollama for local model execution, with a practical path to swapping models
• Local embedding services built on IBM Granite English retrieval models

Crucially, the validated configuration runs on CPU. No GPU was required for the deployed evaluation workflow, which lowers the infrastructure barrier considerably for on-premises adoption.

Retrieval Tuned for Engineering Content

The heart of the deployment is a hybrid retrieval path designed around how technical documentation actually behaves. Memuron combines three signals in a single search:

1. Semantic matching for natural-language questions
2. Lexical signals for domain terminology
3. Exact-identifier handling for filenames, part references, and technical codes

Memuron also searches small, source-linked evidence units and preserves document and section provenance, so every piece of retrieved context can be traced back to its source.

Alongside the stack, AlphaNimble handed over installation guidance, configuration references, health checks, model-download steps, and evaluation tooling, so the ASM team could set up and test the system on their own.

The Results: Validated by the Client

ASM ran three independent evaluation runs in its own environment. The numbers speak clearly:

A few things stand out.

Complete evidence discovery. The system surfaced the required source material for every evaluated question. For engineering use cases, this matters more than almost anything else. An answer can only be as good as the evidence behind it.

Strong grounding. With 93% faithfulness, generated answers closely followed the context Memuron retrieved rather than drifting into unsupported claims.

Consistency. All three runs landed within a tight band of roughly 1.25 points, pointing to stable, predictable system behavior rather than a lucky result.

Reading the Scores Honestly

A good evaluation tells you what is working and where to go next. ASM's results did both.

Since recall is already perfect, the opportunity lies downstream. Context relevance came in at 0.69, meaning the retrieved context often included more material than a question needed. Answer correctness sat at 0.77, indicating the generator did not always convert good evidence into the expected answer.

In other words: Memuron is finding the right material. The next gains come from sending the model a tighter, cleaner evidence set and giving it clearer instructions on how to use it.

The Roadmap: From 84.96 to 89+

AlphaNimble has proposed a structured improvement plan with a minimum target of 89.21, a 5% lift over the baseline, while protecting the 100% evidence-discovery result at every step.

Step 1: Establish a controlled baseline. Freeze the corpus, embedding model, retrieval configuration, evaluation set, judge model, and scoring version, and log detailed per-query diagnostics. Then tie every future score change to a specific configuration change.

Step 2: Trim context before generation. Keep the Granite model and current candidate retrieval, but pass only the top three chunks to the generator, with a configurable token cap. This directly targets excess-context cases, aiming for at least 0.85 context relevance.

Step 3: Rerank and deduplicate. Add a lightweight reranking stage that favors the most direct sections, heading matches, and source diversity, while suppressing near-duplicate chunks. Exact part numbers, filenames, and codes keep a priority override.

Step 4: Make generation evidence-bound. Instruct the model to answer directly, rely only on supplied evidence, cite its sources, and clearly state when the corpus does not contain an answer. Targets: answer correctness of at least 0.85 and faithfulness of at least 0.95.

Scaling Ingestion Without Losing Quality

As ASM broadens its corpus, ingestion quality becomes the next frontier. The recommended workstream includes a source manifest with checksums and versioning; extraction validation that quarantines empty or failed documents; structure-aware chunking by headings, procedures, and tables; normalized technical metadata for part numbers and revisions; duplicate and version control; and golden-set acceptance tests before promoting any corpus update.

The rule is simple: a new configuration ships only if it improves component scores on the frozen evaluation set and keeps evidence discovery at 1.00.

Why This Matters

This engagement shows that trustworthy, grounded AI over technical documentation does not require giving up control of your data or investing in GPU infrastructure. ASM now has a validated on-premises foundation: its corpus stays in-house, the retrieval stack runs on CPU, and local Ollama models can be configured for downstream use.

For AlphaNimble, it reinforces what Memuron was built to do: find the right evidence, keep it traceable, and give engineering teams answers they can trust.

ASM Technologies reported performance figures and deployment details after three evaluation runs in August and September 2026.