Built. Shipped. Measured.

AI-driven analysis of multi-spectral thermal drone imagery to detect hidden subsurface deterioration in critical infrastructure — replacing expensive, dangerous manual inspection with safer, faster, more comprehensive coverage.

Soledad's engineers deployed multi-spectral drones with near-infrared payloads to capture rich thermal gradient data across the full arc of the day. A diffusion-model architecture, trained on those temperature datasets, distinguishes thermal patterns indicative of healthy material from those suggesting hidden degradation — generating GeoJSON mask overlays that spatially map areas of concern onto each structure's geographic footprint.

The architecture lives on Azure-hosted edge infrastructure rather than the drone, with the fleet acting as a data-collection instrument. The result is repeatable, scalable, and safe subsurface degradation analysis at portfolio scale — letting maintenance teams prioritize intervention before deterioration reaches a critical threshold.

A continuous-monitoring drone platform with on-board AI inference — detecting gas leaks, oil spills, and emerging security threats across hostile oil field terrain with 24/7 autonomous coverage and zero personnel in the hazard zone.

Oil field environments are uniquely hostile to conventional inspection: gas is invisible, contamination spans irregular surfaces, and security threats can emerge anywhere across remote sites. Soledad's Solutions Engineers architected a platform around three coordinated detection modalities — multi-spectral gas leak identification, oil-spill classification across varied land surface types, and adaptive reconnaissance that redirects flight paths mid-mission in response to flagged events.

The team deliberately chose on-drone inference over edge-streamed inference. Field connectivity is unreliable, and threat investigation cannot wait on round-trip latency. Embedded AI was deployed directly to drone compute, with a forward docking station serving as the field operations hub for charging, image offload, and over-the-air model updates — eliminating the need to return drones to a central facility between missions.

Replaced a fatigue-prone, eight-hour-per-shift manual X-ray inspection process with a real-time AI anomaly detection platform deployed on AWS Outposts — delivering low-latency inferences directly to floor technicians while keeping humans in control of final decisions.

The legacy workflow asked a single technician to monitor a continuous X-ray video feed for an entire shift, making real-time routing decisions on every detected anomaly. The cognitive load was unsustainable, and the process created a single point of failure: missed detections meant defective units advanced downstream; false positives meant unnecessary rework. Soledad's Solutions Engineers mapped the existing workflow, anomaly taxonomy, remediation routing logic, and latency budget — translating domain expertise into precise technical requirements.

Our Data Science team trained a YOLO-architecture model on labeled historical footage, balancing detection accuracy with the inference speed a streaming video pipeline demands. Full-Stack Engineering wrapped the model in a purpose-built operator dashboard with confidence-graded anomaly flags, recommended routing actions, and a feed history log. Systems Architecture ruled out cloud round-trips and led the deployment of AWS Outposts on the production floor — delivering cloud-consistent tooling with on-prem latency.

A pre-LLM-era deployment of Word2Vec and Bag-of-Words sentiment models into a live enterprise chat support environment — among the earliest production NLP integrations in customer service, and a direct contributor to the research lineage that became modern large language models.

When text-based support began replacing voice in enterprise customer service, the paralinguistic cues experienced agents had relied on — tone, pace, hesitation — disappeared from the channel. There was no tooling to interpret the emotional or intentional content of a live chat, and escalation risk was visible only after a conversation had already closed. Soledad's engineers were brought in to close that gap. We mapped the agent and supervisor workflows in-situ, defined the requirements for unobtrusive in-interface sentiment signals, and built a platform that produced live sentiment scoring at the message level.

The Data Science work combined Bag-of-Words feature extraction with Word2Vec embeddings to characterise the trajectory of an ongoing conversation — identifying interactions trending toward resolution versus deterioration in near real time. Front-end engineers translated those signals into an in-chat sentiment indicator and a supervisor queue view that flagged at-risk interactions. The NLP techniques operationalised here were research-grade at the time. The applied work done on this class of problem fed directly into the lineage that produced today's transformer architectures and large language models. Soledad was not observing that transition. We were participants in it.

A reference architecture that reassigns AI workloads by task complexity — routing high-volume code generation, embedding, and retrieval to locally deployed Ollama models on Docker, and reserving frontier agents like Claude and Gemini for genuine reasoning. Combined with a local Elasticsearch RAG pipeline, the design typically reduces enterprise AI operating spend by 60–85%.

Most enterprises adopting AI-assisted development route every developer query — one-line functions, test scaffolds, documentation stubs, multi-system architectural decisions — to the same frontier model at the same per-token price. At scale that is the equivalent of dispatching a senior architect to fill out a change-of-address form. Soledad's ML Architects redistribute the work: a local tier built on Ollama embedding and code-generation models, sharing a Dockerised Elasticsearch hybrid index (kNN + BM25), handles the high-volume retrieval and generation. A cloud tier of frontier agents handles research, planning, validation, and orchestration — where their cost is justified by capability.

The accompanying data-flow diagram makes the boundary explicit. Cloud or external AI agents drive the planning and validation loop; local Docker infrastructure runs the indexing, generation, and Playwright test execution. Outputs — documentation, code standards, generated code, validated merges, and a rework queue — feed back into the vector store so each subsequent cloud call benefits from accumulated organisational context. For a 100-engineer organisation, the model predicts annual AI spend dropping from approximately $995,000 to $250,000 — a 75% reduction — while improving code quality, maintaining data sovereignty, and producing a proprietary knowledge asset that compounds over time.

Some of Soledad's clearest engineering principles came not from what worked, but from what didn't. This retrospective documents an ad-buying engagement where the validation layer the specification asked for passed every test — and the validation layer it did not ask for would have prevented the platform-ban cascade that ended the client.

Soledad's engineers were embedded inside an ad-buying organisation that placed creative across Facebook, Google, and adjacent networks. Two priorities dominated the roadmap: faster iteration on live deployments, and a bid-range recommender for the digital-marketing real estate the client cared about most. We built both. The full-stack application gave buyers in-app text and image insertion with direct push to the integrated networks. The collaborative developer–end-user workflow matched the operational reality of a buying desk. Iteration speed went up. Bid efficiency improved. By the engineering metrics we had agreed to, the engagement was successful.

Output validation enforced exactly what the specification asked for: character counts, image dimensions, required fields. What it did not include — and was not asked for — was qualitative review of the content itself against the published acceptable-content policies of the destination platforms. There was no image analysis pass and no NLP review of body text against those standards. The application would happily accept a perfectly-sized, perfectly-lettered creative whose imagery or copy crossed a policy line, and dispatch it to the network. Mis-steps accumulated, platform enforcement followed, and the ban cascade cut off the inventory the business depended on. The company will remain unnamed. RIP.

Experiences like this one have shaped how Soledad's product and engineering leadership scope, review, and ship new work. Qualitative compliance is part of the definition of done. Where output lands on a third-party platform, qualitative review against that platform's published policy is part of the validation layer — not a stretch goal. Any generative system we build ships with a content-policy evaluation step before output reaches a destination the customer is held accountable for. Third-party platforms whose rules can shut a customer down are treated as governance surfaces with policy versions tracked. Engagements involving PII, financial, health, or biometric data carry an explicit data-handling contract that names the validation and review steps that apply — and those steps are not negotiated away to make a deadline.